EP4677336A1

EP4677336A1 - Systems and methods for soil classification based on hyperspectral imaging

Info

Publication number: EP4677336A1
Application number: EP24707528.6A
Authority: EP
Inventors: Neda SEPASIAN; Diego CARVALHO; Joseph Turner
Original assignee: FNV IP BV
Current assignee: FNV IP BV
Priority date: 2023-03-03
Filing date: 2024-02-26
Publication date: 2026-01-14
Also published as: NL2034265B1; CN120641732A; CL2025002553A1; WO2024184107A1

Abstract

The disclosure relates to a method for predicting soil characteristic information, the method comprising: receiving data representative of a hyperspectral image of a soil sample; determining a hyperspectral reflectance profile for the soil sample based on the received data, wherein the hyperspectral reflectance profile comprises reflectance intensity information; determining one or more soil characteristics based on the determined hyperspectral reflectance profile of the soil sample, wherein the one or more soil characteristics include one or more of: soil-type classification information for the soil sample; chemical composition information for the soil sample; particle size information for the soil sample A machine learning method for classifying soil samples is also disclosed, along with a training model, and associated systems for carrying out methods according to the disclosure. Unlocking insights from Geo- Data, the present invention further relates to improvements in sustainability and environmental developments: together we create a safe and liveable world.

Description

SYSTEMS AND METHODS FOR SOIL CLASSIFICATION BASED ON HYPERSPECTRAL IMAGING

TECHNICAL FIELD

[0001] The disclosure relates to methods and systems for predicting soil characteristic information based on data representative of a hyperspectral image of a soil sample. The disclosure also provides a method of training a machine learning model to identify a characteristic of a soil sample, a method of obtaining an estimate of soil classification using a machine learning model, and systems and computer readable media configured to do the same. Unlocking insights from Geo-Data, the present invention further relates to improvements in sustainability and environmental developments: together we create a safe and liveable world.

BACKGROUND

[0002] There is a general and ongoing need for systems and methods for determining or estimating the characteristics of collected or observed soil samples. One of the goals of soil classification is to quantify or estimate the proportion of soil types that constitute a soil sample. Such information is useful because it may be used as a predictor or indicator of the physical properties of the soil, which are relevant for the planning and design of structures that interact with the soil environment. For example, the physical properties of soil located at a site earmarked for construction may impose different requirements on the foundations of a structure.

[0003] In situ soil classification may be performed in geological surveys. In such surveys, specialists in the field of geology and soil mechanics may observe the visual aspect of the soil, including, for example, the soil colour and granularity. Since this preliminary analysis is dependent on human perception of the soil, consensus among peers is generally needed to provide an unbiased assessment of the soil classification. In addition to increasing logistical challenges gathering sufficient expertise to assess the soil in situ, analysis based on human perception is prone to unquantifiable errors and bias, and lacks reproducibility.

[0004] Moreover, soil classification techniques may be impacted by the dryness (or otherwise) of the sample, since the visual appearance of a sample may vary greatly depending on the saturation and moisture content of the sample, in situ soil classification is vulnerable to variability due to changing environmental conditions at the sample location.

[0005] Aside from in situ soil classification, soil samples may be collected and returned to a laboratory test environment for analysis. For example, soil samples may be evaluated in a controlled environment, and compared to reference samples available in a laboratory. Samples may also be prepared under controlled conditions (e.g. samples may be dried in an oven) before analysis. Soil attributes such as chemical composition and pH may be analysed.

[0006] A need remains for improved soil classification techniques that can accurately and reliably classify soil types, with improved reproducibility and without the logistical challenges posed by current standard techniques. SUMMARY

[0007] The present disclosure provides systems and methods for improved soil classification.

[0008] According to a first aspect of the disclosure, there is provided a method for predicting soil characteristic information, the method comprising: receiving data representative of a hyperspectral image of a soil sample; determining a hyperspectral reflectance profile for the soil sample based on the received data, wherein the hyperspectral reflectance profile comprises reflectance intensity information; determining one or more soil characteristics based on the determined hyperspectral reflectance profile of the soil sample, wherein the one or more soil characteristics include one or more of: soil-type classification information for the soil sample; chemical composition information for the soil sample; particle size information for the soil sample. The method can be a computer-implemented method. The reflectance profile can comprise a reflectance intensity spectrum across the hyperspectral range of interest, for example in the VNIR range. Although the examples described in this document comprise profiles measured in the VNIR range, the present disclosure is not so limited.

[0009] Optionally, determining one or more soil characteristics based on the determined hyperspectral reflectance profile of the soil sample comprises comparing the determined hyperspectral reflectance profile to one or more known reflectance spectra and/or profiles.

The hyperspectral reflectance profile may comprise reflectance intensity information for wavelengths in the range of: 400nm to 2000nm; 400nm to 1000nm, or between 400nm to 550nm and between 850nm to 1000nm. In at least one example, the spectral profile comprises reflectance intensity information for wavelengths below approximately 600nm and above approximately 800nm, optionally below 550nm and above 850nm.

[0010] The hyperspectral image of a soil sample can be a snap-shot hyperspectral image.

In some embodiments, the hyperspectral image can comprise a spectral sampling interval of approximately 20nm or less, approximately 15nm or less, approximately 10nm or less, or approximately 5nm or less. The sample interval may be constant and extend over substantially the whole captured wavelength range. Alternatively, the spectral interval may vary across the wavelength range. The sample interval may be less than 20nm, optionally less than 10nm, optionally less than 5nm.

[0011] The method can further comprise: identifying one or more regions of interest, ROIs, in the hyperspectral image; determining the hyperspectral reflectance profile for the one or more ROIs.

[0012] A plurality of ROIs can be identified, and the determining the hyperspectral reflectance profile of the ROIs can comprise determining a mean hyperspectral reflectance profile for the plurality of ROIs. [0013] The hyperspectral image is captured at a distance D from the soil sample, and wherein D is approximately 100m or less, optionally approximately 25m or less, optionally approximately 10m or less, optionally approximately 5m or less, optionally approximately 1 m or less.

[0014] The method can further comprise the step of capturing the hyperspectral image.

[0015] The image can be captured of a soil sample in situ or wherein the image is captured of a soil sample in a test environment. [0016] The soil-type classification information can comprise a soil classification label, such as clay, silt, and/or sand.

[0017] The step of determining one or more soil characteristics based on the determined hyperspectral reflectance profile of the soil sample can comprise using a machine learning model to predict the soil characteristics.

[0018] In another aspect of the disclosure, there is provided a system for capturing hyperspectral reflectance information of a soil sample, the system comprising: a remotely operated vehicle, ROV, comprising one or more hyperspectral cameras configured to capture a hyperspectral image of a soil sample; wherein the one or more hyperspectral cameras is configured to detect a hyperspectral reflectance profile comprising reflectance intensity information for wavelengths in the range of: 400nm to 2000nm; 400nm to 1000nm, or between 400nm to 550nm and between 850nm to 1000nm [0019] The ROV can further comprise a light source, and, optionally a calibration surface.

[0020] The system can further comprise a communication module configured to transmit hyperspectral image data to a remote storage module.

[0021] The system can further comprise one or more processors configured to carry out the steps of any of methods described herein.

[0022] According to another aspect, there is provided a method of training a machine learning model to identify a characteristic of a soil sample, the method comprising: obtaining a hyperspectral reflectance profile associated with a soil sample; obtaining a target characteristic associated with the soil sample; providing the hyperspectral reflectance profile to a machine learning model to obtain an output of the machine learning model, the output of the machine learning model comprising a determined characteristic of the soil sample; and adjusting parameters of the machine learning model to reduce an error between the determined characteristic of the soil sample and the target characteristic of the soil sample.

[0023] In this manner, the machine learning model is trained to classify hyperspectral reflectance spectra associated with images of soil samples, so as to output (predict) a particular soil characteristic of interest. For example, the machine learning model may be trained to output a label indicating the primary constituent of the soil sample, for example clay, sand or silt. The accuracy of the output is assessed against a ground truth target label, and the model parameters are adjusted to improve the accuracy of the prediction. Through iterative training and adjustment of the model parameters in this manner, the machine learning model becomes better at the classification task and can then be used to implement the soil classification methods described above and in more detail herein. Hence, this training method facilitates an automated, reliable mechanism for soil sample classification that addresses the shortcomings with existing classification and analysis methods outlined above.

[0024] The hyperspectral reflectance profile may advantageously comprise reflectance data for wavelengths between 400 and 1000nm. This spectra of wavelengths can be easily obtained using relatively cheap cameras. Optionally, the range of wavelengths analysed may be enlarged to include wavelengths further in the infra-red and UV ranges which may assist in classification, although this typically requires more complex cameras. [0025] As noted above, the output of the machine learning model is a determined (predicted) characteristic of the soil sample, which can be compared to a (ground truth) target characteristic to assist in training. As noted above, one example of a soil characteristic comprises a label indicating the primary (i.e. most prevalent) constituent of a particular soil sample, typically sand, silt or clay. That said, the training method can focus on any suitable soil characteristic that is of interest. For example, the determined and target characteristics of the soil sample may comprise one or more of: soil-type classification information for the soil sample; chemical composition information for the soil sample; and/or particle size information for the soil sample. Analysis of each of these characteristics can provide useful insight into the structure, composition and characteristics of the soil sample.

[0026] The present inventors have identified that the effectiveness of the classifier training process described above may be improved through use of a pre-training process, carried out prior to the primary classifier training. In particular, implementation of reconstruction pre-training for the machine learning model (or one or more components thereof) prior to use of the machine learning model for classification training can assist in the subsequent classification training. Accordingly, the method may comprise, prior to providing the hyperspectral reflectance profile to the machine learning model, performing a pretraining process on one or more components of the machine learning model. The pre-training process may advantageously comprise: obtaining a target high resolution hyperspectral reflectance profile; and obtaining a low resolution hyperspectral reflectance profile. The low resolution hyperspectral reflectance profile may comprise a down sampled or binned version of the target high resolution hyperspectral reflectance profile. The method further comprises: providing the low resolution hyperspectral reflectance profile to the one or more components of the machine learning model to obtain an output of the one or more components of the machine learning model, the output comprising a determined high resolution hyperspectral reflectance profile; and adjusting parameters of the one or more components of the machine learning model to reduce an error between the determined high resolution hyperspectral reflectance profile and the target high resolution hyperspectral reflectance profile.

[0027] In this manner, the machine learning model (or one or more components thereof) is pre-trained to reconstruct high resolution hyperspectral reflectance spectra from low resolution hyperspectral reflectance spectra. Through this pre-training, the pre-trained component(s) of the machine learning model effectively learn to “understand” the features of hyperspectral reflectance spectra. This acquired understanding then assists the machine learning model when it is subsequently subjected to classifier training relating to soil sample spectra. A key advantage of the reconstruction based pre-training is that any hyperspectral reflectance spectra can be used during this pre-training, including unlabelled hyperspectral reflectance spectra that are not related to images of soil samples. Such data is readily available in large volumes, in contrast to the more specific labelled hyperspectral soil spectra needed for the full classifier training. Hence, the pre-training addresses a problem that hyperspectral images and associated spectra of soil samples are scarce, which can limit the possibility for effective training using only classifier training with labelled soil sample input data.

[0028] It will be appreciated that, where pre-training is used, it is not necessary that all components of the machine learning model undergo pre-training. Rather, in some implementations only one or some components of the model are pre-trained. More generally, in other words, a different model architecture can be used during pre-training and full classifier training.

[0029] In one implementation, the machine learning model comprises an encoder. If pre-training is used, the encoder can be pre-trained, for example in an encoder-decoder architecture. After pre-training, the pre-trained encoder is returned to the original machine learning model architecture and used within the classifier training of the model. The use of an encoder in this manner has been found to result in a particularly effective classifier for hyperspectral soil sample data.

[0030] In one implementation, the machine learning model may comprise an implicit neural representation network, optionally a sinus implicit neural representation network (SIREN). The present inventors have identified that implicit neural representation networks (and in particularly SIRENs) are particularly well suited to hyperspectral reconstruction and classification tasks. Advantageously, the machine learning model may comprise a residual network (ResNet) which has been found by the present inventors to be particularly well suited to hyperspectral reconstruction and classification tasks.

[0031] The machine learning model may be further trained with hyperspectral image data comprising one more labels indicative of sample wetness. This can assist the machine learning model in understanding the impact of soil sample wetness on the hyperspectral reflectance profile.

[0032] According to yet another aspect, there is provided a method of obtaining an estimate of soil classification using a machine learning model trained in accordance with any of the machine learning methods described above, wherein the method comprises: obtaining input data comprising data representative of a hyperspectral image of a soil sample; applying the input data to the machine learning model to obtain an output of the machine learning as the estimate.

[0033] According to yet another aspect of the disclosure, there is provided a system comprising: one or more processors; and one or more memories having stored thereon computer readable instructions configured to cause the one or more processors to perform operations comprising the steps of any of the methods described herein.

[0034] The system can comprise one or more sensor systems, wherein the one or more sensor systems optionally comprises a hyperspectral camera.

[0035] According to another aspect of the disclosure, there is provided one or more computer readable media comprising instructions, that, when executed by one or more data processing apparatus, cause the one or more data processing apparatus to perform operations comprising the steps of any of the methods described herein.

[0036] According to another aspect of the disclosure, there is provided a machine learning model stored on one or more computer readable media, wherein the model has been trained according to one or more of the training methods described herein.

BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The disclosure will be further described with reference to illustrative embodiments and in connection with the following drawings, in which:

Figures 1 a to 1 c show three different soil samples; Figures 2a to 2c shows the hyperspectral reflectance profile for three different soil samples;

Figure 3 shows an exemplary system for capturing a hyperspectral image and determining soil classification information based on the captured image according to the disclosure;

Figure 4 shows a schematic of an exemplary remote operated vehicle comprising a hyperspectral image capture device according to the disclosure;

Figure 5 shows a flowchart of a method of soil classification according to the disclosure;

Figure 6 shows, in schematic form, a method of training a machine learning model to identify a characteristic of a soil sample according to the disclosure;

Figure 7 shows, in schematic form, a particular implementation of the method from Figure 6; Figure 8 shows, in schematic form, a pre-training method according to the disclosure;

Figures 9a to 9c show example inputs and outputs for the machine learning model according to the disclosure;

Figures 10a and 10b show an exemplary network architecture that may be used during pretraining according to the disclosure;

Figure 11 shows a computer system for carrying out the various methods of the disclosure.

DETAILED DESCRIPTION OF THE DRAWINGS

[0038] The following detailed description is merely exemplary in nature and is not intended to limit the application and its uses. Furthermore, there is no intention to be bound by any expressed or implied theory presented in the preceding technical field, background, brief summary or the following detailed description. As used herein, the term ‘module’ refers to any hardware, software, firmware, electronic control component, processing logic, and/or processor device, individually or in any combination, including without limitation: application specific integrated circuit (ASIC), an electronic circuit, a processor (shared, dedicated, or group) and memory that executes one or more software or firmware programs, a combinational logic circuit, and/or other suitable components that provide the described functionality.

[0039] Embodiments of the present disclosure may be described herein in terms of functional and/or logical block components and various processing steps. It should be appreciated that such block components may be realised by any number of hardware, software, and/or firmware components configured to perform the specified functions. For example, an example embodiment of the present disclosure may employ various integrated circuit components, e.g. memory elements, digital signal processing elements, logic elements, look-up tables, or the like, which may carry out a variety of functions under the control of one or more microprocessors or other control devices. In addition, those skilled in the art will appreciate that embodiments of the present disclosure may be practised in conjunction with any number of systems, and that the systems described herein are merely exemplary embodiments of the present disclosure.

[0040] For the sake of brevity, conventional techniques compared to signal processing, data transmission, signalling, control and other functional aspects ofthe systems (and the individual operating components of the systems) may not be described in detail herein. Furthermore, the connecting lines shown in the various figures contained herein are intended to represent example functional relationships and/or physical couplings between the various elements. It should be noted that many alternative or additional functional relationships or physical connection may be present in an embodiment of the present disclosure.

[0041] Systems and methods described herein relate generally to systems and methods for predicting soil characteristic information, such as soil type, based on received hyperspectral data. The received data may be a hyperspectral image of a soil sample. Determining soil type (e.g. clay, silt, sand) allows soil samples to be classified into a labelled group that can provide important insights into the physical properties of the soil, which may be predictive of the behavior of the soil during and after construction of a structure.

[0042] The present disclosure generally provides methods and systems for determining a soil type based on an observed sample. The systems and methods described herein provide automated (or semiautomated) determination of soil type. The automated systems are configured to receive data representative of a hyperspectral image of a soil sample and to determine, by comparison of the hyperspectral reflectance profile to reference spectra and/or values, a soil type for the sample based on the comparison.

[0043] The systems and methods described herein may also make use of a machine learning model to determine the soil type based on received hyperspectral information, as will become apparent from the description of the detailed embodiments that follow.

[0044] Figures 1 a to 1 c each show a soil sample 100a, 100b, 100c, with the magnified portions B showing, in schematic form, a visually different appearance for each of the samples. The visually different characteristics of the soil samples may include colour, particle shape, particle size, and a combination thereof. Aside from the differing visual attributes of the soil samples, each of the different samples may have a different hyperspectral profile.

[0045] Figures 2a to 2c each show the hyperspectral reflectance profile of a soil sample. Figures 2a to 2c show the reflectance intensity for different wavelengths of incident light on a soil sample. In example shown in Figures 2a to 2c, the reflectance profile spans a wavelength range of approximately 400nm to approximately 1000nm. In other words, the hyperspectral reflectance profile is in the visual and near infra-red range (VNIR). However, the present disclosure is not limited to the wavelength range shown in Figures 2a to 2c. For example, the range over which reflectance may be measured can be extended to the mid-infra-red or beyond. In at least one implementation, a reflectance profile can be collected over a wavelength range of 400nm to 2500nm, optionally a wavelength range of 400nm to 200nm, optionally a wavelength range of 400 to 1500nm, and optionally a wavelength range of 400 to 100nm.

[0046] Figure 2a shows a hyperspectral reflectance profile 200a for a soil sample classified as clay. As shown in Figure 2a, the sample shows an identifiable primary reflectance peak, approximately 0.36, at a wavelength of approximately 960nm. As shown in Figure 2a, the reflectance profile 200a for the clay sample shows a (relatively) small linear increase in the reflectance from approximately 400nm until the approximated wavelength of 900nm, when a steep increase in reflectance intensity is observed. The reflectance peaks at a wavelength of approximately 960nm and falls sharply until approximately 1000nm. This profile, including the shape of the profile, the gradient of the reflectance across all or parts of the spectrum, the number of and/or location (within the wavelength range) of identified peaks, may be characteristic of a particular soil type, in this case clay.

[0047] Figure 2b shows a hyperspectral reflectance profile 200b for a soil sample classified as sand. As shown in Figure 2b, the sample shows an identifiable reflectance peak, approximately 0.56, at a wavelength of 950nm. As shown in Figure 2b, the profile 200b does not present linear growth (as in Figure 2a). Rather, there is steeper reflectance growth between 400nm and 500nm. Between 600nm and 900nm there is substantially linear growth, before the reflectance increases steeply from approximately 900nm. The peak at 950nm is observed, before the reflectance again falls steeply towards 1000nm. This profile, including the shape of the profile, the gradient of the reflectance across all or parts of the spectrum, the number of and/or location (within the wavelength range) of identified peaks, may be characteristic of a particular soil type, in this case sand.

[0048] Figure 2c shows a hyperspectral reflectance profile 200c for a soil sample classified as silt. As shown in Figure 2b, the sample shows an identifiable reflectance peak, approximately 0.52, at approximately 950nm. A secondary peak, at approximately 880nm can be observed, with a reflectance of approximately 0.44. Reflectance is generally stable between 400nm and 450nm, with substantially linear growth between 450nm and 850nm. Reflectance falls between the first and second peaks (at approximately 850nm and 950nm), and again drops sharply from the primary peak at approximately 950nm towards 1000nm.. This profile, including the shape of the profile, the gradient of the reflectance across all or parts of the spectrum, the number of and/or location (within the wavelength range) of identified peaks, may be characteristic of a particular soil type, in this case silt.

[0049] In the spectra shown in Figures 2a to 2c, the reflectance intensity was measured at a spectral sampling interval of 4nm. However, it will be appreciated that other spectral sampling ranges may be selected. The spectral interval is preferably less than 20nm, preferably less than 15nm, and preferably less than 10nm, and preferably less than 5nm.

[0050] The spectral interval in Figures 2a to 2c is also constant across the measured range (between 400nm and 100nm). However, it will be appreciated that the spectral interval may vary across the measured range. For example, the spectral sampling interval may be greater in spectral regions that provide few characteristic peaks relevant to soil type than others. For example, as can be seen in the spectra from Figures 2a to 2c, identifiable spectral profile features can be found in the wavelength range of less than approximately 600nm and at wavelengths of greater than approximately 800nm for the three exemplary soil types shown. Therefore, the sampling frequency may be higher (resulting in a smaller sampling interval) in the spectral regions of interest, e.g. below approximately 600nm or approximately 550nm, and above 800 or 850nm, than in regions of lesser interest, such as the region between 600nm or 800nm or between 550nm and 850nm.

[0051] It will be appreciated in the discussion above that hyperspectral reflectance profiles for a soil sample may be indicative of soil type. In the examples provided in Figures 2a to 2c, the maximum reflectance is indicated in each profile on the y-axis. However, it will be appreciated that the maximum reflectance value for a sample is less relevant to the characterisation of the soil type than the location of and relative intensity between peaks. This approach to identification of soil samples may be particularly advantageous because the profile is largely unaffected by the wetness of the sample. The inventors of the present application have observed that although the visual appearance of a sample changes with the degree of water saturation of the sample (e.g. the wetter the sample the darker it appears), and although the intensity of reflected radiation varies with wetness, the shape of the profile and the location the peaks does not. Systems and methods according to the disclosure may therefore provide additional improvements over traditional methods of determining soil type, which may require additional steps to

[0052] Turning to Figure 3, an exemplary system for determining a hyperspectral profile will now be described. As shown in Figure 3, system 300 comprises a hyperspectral image capture device 304. The device may be configured to capture a snap-shot hyperspectral image. The hyperspectral image capture device can comprise a hyperspectral camera, such as the VNIR 4250 available from Hinalea Imaging. The image capture device is configured to capture an image in which the intensity of (reflected) electromagnetic radiation is recorded for a plurality of wavelengths for each image pixel, within the visible range and beyond. The hyperspectral image therefore provides spatial information (as in a conventional RBG image) in addition to reflectance intensity information for the plurality of wavelengths. As mentioned previously, the range of interest may comprise the VNIR range. The hyperspecrtal image capture device may also be configured to measure reflectance intensity Short Wave Infrared range (SWIR).

[0053] Although in the example described above with reference to Figure 3, a single hyperspectral image capture device is shown, the present disclosure also includes embodiments in which a plurality of hyperspectral image capture devices are used, configured to determine hyperspectral images for distinct (and optionally overlapping) wavelength ranges.

[0054] As shown in Figure 3, the image capture device 304 (or the plurality of image capture devices) is in operative communication with a processor 306 and a data store 308. The processor 306 may be configured to control the image capture device. The data store is configured to store the captured image data. The captured image data may comprise a data cube, representing a 2D array of pixels, with intensity information for the plurality of wavelengths provided by way of multiple channels for each pixel. The data cube representing the hyperspectral image can comprise the information required to determine one or more hyperspectral profiles for the imaged sample, as will be explained in further detail below. The processor of the image capture device may be configured to carry out some or all of the data processing steps described below. Alternatively, the image capture device may be configured to export the image data to an external system for processing, as illustrated in Figure 3.

[0055] In embodiments in which a plurality of image capture devices is used to capture a plurality of overlapping or non-overlapping wavelength ranges, the data can be stored in a plurality of data cubes, or it may be consolidated into a single data cube.

[0056] It will be appreciated that, in the context of the present disclosure, a hyperspectral profile may be determined from a single pixel of a captured hyperspectral image, by determining reflectance intensity for each wavelength interval across the measured wavelength range. Alternatively, a hyperspectral profile can be determined based on a plurality of pixels, for example, by averaging the reflectance intensity for a plurality of selected pixels. The selected pixels may comprise all the pixels with a selected regions of interest (ROI) of the hyperspectral image. A plurality of ROIs may be identified and a spectral profile determined for each ROI, based on, e.g., mean intensity for the pixels of an individual ROI. The hyperspectral profiles of each ROI may then be compared to each other to identify inconsistencies between the spectral profiles. Such a comparison step may be used to exclude the pixels of a non-representative ROI from the determined hyperspectral image for the sample. Alternatively, hyperspectral profiles for multiple ROIs may be determined from a single hyperspectral image to identify mixed-type samples, wherein a first ROI is classifiable as a first soil type and a second ROI is classifiable as a second soil type.

[0057] In yet another example, a hyperspectral profile can be determined based on all pixels of the captured hyperspectral image or on all pixels of a hyperspectral image associated with the sample. In one implementation, the image data may be processed to exclude data representative of any regions of pixels having reflectance exceeding a threshold value as representative of the background or calibration surface). The hyperspectral profile information may then be determined as a mean of the reflectance of all pixels not deemed to form part of the background.

[0058] In yet further embodiments, individual pixels exceeding threshold values for reflectance may be excluded as non-representative of the sample.

[0059] In the preceding paragraphs, exemplary image processing steps have been described, in which one or more hyperspectral profiles representative of a soil sample can be determined. It will be appreciated that the system 300 may be configured to carry out some or all of these image processing steps. Alternatively, the system 300 may be configured to store the hyperspectral image data and export the raw data to an external system for processing.

[0060] As shown in Figure 3, the system 300 may be a self-contained capture system, configured to capture a hyperspectral image and store the image data for subsequent processing in an external system. In such embodiments, the system 300 can comprise a wired or wireless connection to an external data processing system.

[0061] In the embodiment shown in Figure 3, the system 300 is in operative communication with a network device 310. The network device 310 is in operative communication with a first terminal 314, which comprises a personal computer, and a second terminal 312, which comprises a mobile device. The end user terminals may further comprise additional processing means to process the hyperspectral image data to determine a soil type based on the captured image.

[0062] Although the system 300 described above is configured for wireless communication via a network device 310, it will be appreciated that one or more of the connections may be wired.

[0063] Although not shown, the system 300 may further comprise or be configured for use with a calibration surface. The calibration surface can comprise a surface with known hyperspectral reflectance properties, such as a bright white, matte plate. The calibration surface may be configured to be within the field of view of the camera 304 when an image of a sample is captured.

[0064] The processor 306 and/or a processor in communication with the system 300 (such as at one of the terminal devices) may be configured to determine one or more regions of interest (ROIs) for the soil sample. An ROI may be, for example, towards an edge of the soil sample, adjacent to a region of a calibration surface in the field of view of the camera 304. Identification of an ROI by the image capture device or a processor in communication with system 300 may be advantageous since the volume of data required for upload to a further system for additional processing steps may be reduced.

[0065] Although ROIs may be determined automatically by a processor, it will be appreciated that a human operator may identify one or more ROIs from a hyperspectral image. It will also be appreciated that an ROI need not be a sub-region of the image, but may rather be the whole image of the soil sample. However, selecting one or more sub-regions of a hyperspectral image as an ROI may be advantageous in terms of decreasing volume of data for processing, and for selecting ROIs that are representative of the sample, avoiding anomalous material not representative of the wider sample.

[0066] Turning now to Figure 4, in at least one exemplary embodiment, an image capture device is provided on a remotely operated vehicle (ROV). The ROV can comprise a land-borne ROV, an airborne ROV (e.g. an unmanned aerial vehicle or UAV) or surface-water or submersible ROV. The remotely operated vehicle 400 comprises a body 402, which comprises a hyperspectral image capture device 404, a processor 406 and a storage device 408. The ROV may further comprise a light source 410, for illuminating a sample, and/or a calibration surface 412 having known hyperspectral reflectance properties. One or more of the light source 410 and the calibration surface 412 may be mounted on an articulatable arm, for example a robotic arm, that allows repositioning of the light source and the calibration surface relative to a sample.

[0067] The ROV 400 can also comprises propulsion means 414, configured to move the ROV. For example, propulsion means 414 may take the form of one or more propellers for a submersible ROV. Propulsion means 414 may take the form of one or more rotary blades for an unmanned aerial vehicle. [0068] The ROV may be configured to capture a hyperspectral image of a soil sample, in situ, from a distance of less than 100m, optionally less than 50m, optionally less than 25m, optionally less than 10m, optionally less than 5m and optionally approximately 1 m or less.

[0069] In the preceding paragraphs, and with reference to Figures 1 to 4, exemplary systems have been described for capturing one or more hyperspectral images of soil samples to be classified.

Turning now to Figure 5, a method for classifying a soil sample will now be described. As shown in Figure 5, the method comprises an optional preparatory step, 500, in which a hyperspectral image of a soil sample is captured. At step 502, the method comprises receiving data representative of a hyperspectral image of a soil sample. At step 504, a hyperspectral reflectance profile is determined for the soil sample based on the received data, wherein the hyperspectral reflectance profile comprises reflectance intensity data for a plurality of wavelengths. At steps 506 and 508, one or more soil characteristics is determined based on the hyperspectral reflectance profile of the soil sample by comparing the determined hyperspectral reflectance profile to one or more known reflectance spectra. The one or more soil characteristics include one or more of a soil-type classification information for the soil sample; chemical composition information for the soil sample; and/or particle size information for the soil sample. The soil-type classification can be a soil classification label, such as clay, silt, and/or sand.

[0070] The hyperspectral image of the soil can be a snap-shot image. The snap shot image may comprise a data cube, in which each pixel provides intensity information for a plurality of wavelengths. [0071] The hyperspectral reflectance profile comprises reflectance data for wavelengths in the range of: 400nm to 2000nm; 400nm to 10OOnm, or between 400nm to 550nm and between 850nm to 10OOnm. [0072] The hyperspectral image comprises a spectral sampling interval of approximately 20nm or less, approximately 15nm or less, approximately 10nm or less, or approximately 5nm or less, and wherein the sampling interval optionally extends across one or more of the ranges of above.

[0073] The method further may further comprise, at step 503, identifying one or more regions of interest, ROIs, in the hyperspectral image, and determining the hyperspectral reflectance profile in the one or more ROIs. The determination of the one or more ROIs may comprise selecting an ROI adjacent to a calibration surface. In such implementations, the method further comprises including a calibration surface or a portion thereof in a field of view of the hyperspectral image capture device when capturing an image of the sample.

[0074] In the event that a plurality of ROIs is identified, the method further comprises determining the mean hyperspectral reflectance profile for the ROIs.

[0075] At step 502, a hyperspectral image is captured using a hyperspectral image capture device. The image capture device may be positioned at a distance D from the soil sample during the image capture step. The distance D is preferably approximately 100m or less, approximately 25m or less, optionally approximately 10m or less, optionally approximately 5m or less, optionally approximately 1 m or less.

[0076] Step 502 may comprise capturing an image of a sample in situ, or a sample in a laboratory setting or test environment.

[0077] The image capture step 502 may also comprise directing a light source towards a soil sample, and optionally, a calibration surface. An optional calibration step can comprise a comparison of the hyperspectral reflectance profile for the calibration surface with a measured hyperspectral reflectance profile for the soil sample. In one exemplary embodiment, calibration may be carried out before each image acquisition. For example, a hyperspectral camera may be arranged such that the field of view of the camera is filled by: the soil sample to be imaged and a calibration surface. In one example, a calibration surface may be arranged such that it fills the field of view of the camera, and the soil sample may be positioned on the calibration surface. A light source may be arranged to illuminate the sample and the calibration surface. The hyperspectral image captured by the camera may therefore provide reflectance intensity for the (regions of interest of) the soil sample and reflectance intensity for the calibration surface. This calibration approach may be particularly suited to hyperspectral cameras configured to separate spectra by light decomposition (such as the Hinalea VNIR 4500). It will be appreciated that the calibration step may be omitted in some embodiments, or other calibration methods may be used.

[0078] In addition to determining a soil type based on the hyperspectral reflectance profile of a soil sample, embodiments of the present disclosure may also identify additional characteristics, using hyperspectral image analysis, of the soil sample. Such additional characteristics may comprise, for example, particle size. Determination of particle size may be an additional useful reference point for predicting the behavioural properties of a soil substrate during engineering and construction projects. [0079] A further useful characteristic of a soil sample that may be determined based on a hyperspectral image nay be information regarding the presence of one or more chemical components in a soil sample. This information may be inferred e.g. based on the identified soil type (e.g. sand) or, additionally or alternatively, it may be determined based on the hyperspectral profile determined from a hyperspectral image.

[0080] As will be appreciated, the method 500 described above and with reference to Figure 5 may further comprise using a machine learning model to predict the soil characteristics. The machine learning model, a training model, and a pre-training model will now be described with reference to Figures 6 to 9.

[0081] Turning now to Figure 6, a method of training a machine learning model to identify a characteristic of a soil sample is shown schematically. This training method can be used within the above-described soil analysis context, in other words a machine learning model trained using the method of Figure 6 can be used to perform the soil sample analysis described above, particularly with reference to Figure 5.

[0082] The method of Figure 6 begins, at step 602, by obtaining input soil characteristic data. This is data that has been obtained experimentally from a set of soil samples and from which some aspect or characteristic of the soil samples (for example their composition) can be determined. As described above, a particularly useful form of input soil characteristic data is a hyperspectral reflectance profile, which indicates reflectance behaviour of a given soil sample across a variety of wavelengths and which can be indicative of material properties (e.g. soil type). Such hyperspectral reflectance spectra can be obtained from hyperspectral images taken by hyperspectral cameras, as described above.

[0083] At step 604, the method proceeds by obtaining target soil characteristic data. Target soil characteristic data comprises labels or indicators relating to a particular characteristic of each sample in the set ofsoil samples from which the input data provided at step 602 has been obtained. For example, the target soil characteristic data may comprise a label indicating the primary constituent of each soil sample (e.g. “clay”, “silt”, “sand” etc.). This target data can be used within the training method as a “ground truth” for each soil sample which the machine learning model is attempting to replicate.

[0084] At step 606, the method trains a machine learning model using the input soil characteristic data provided at step 602 and the target soil characteristic data provided at step 604. In particular, input soil characteristic data for a given soil sample is provided to the machine learning model. The machine learning model then predicts, based on the input data, a characteristic of the soil sample. It is then assessed whether the determined (predicted) soil characteristic matches the actual soil characteristic specified in the respective target soil characteristic data provided at step 604. For example, in one implementation the machine learning network may predict whether the primary constituent in a soil sample is clay, silt or sand. This determination is then assessed against the ground truth label provided for that soil sample to facilitate training.

[0085] Training can be carried out by adjusting model parameters of the machine learning model to reduce an error between the model outputs and the target data, as is known in the art. This process is repeated until a sufficiently well trained model is obtained, as measured by a stopping criterion such as a convergence criterion, a criterion on the error or a pre-set number of epochs. Numerous programming languages and libraries are available to implement this process using a large variety of machine learning models. It will be appreciated by a skilled reader that a suitable model based on considerations such as available data and compute and the kind of data to be processed based on routine considerations can be chosen. The precise details of the training process at step X06 will vary based on these implementation details, as will be appreciated by a skilled reader. The ultimate outcome of the method of Figure 6 is a machine learning model that has been trained to identify soil characteristics based on soil input data such as hyperspectral reflectance profiles of soil samples. A machine learning model trained in this manner can thus be used to implement the above described classifying methods, in particular the method of Figure 5.

[0086] It will be appreciated that the specific type of soil data provided as input and target data at steps 602 and 604 will depend on what soil characteristic is being studied. A particular implementation based on determining soil characteristics from hyperspectral reflectance spectra data is shown in Figure 7.

[0087] The training method of Figure 7 has been developed by the present inventors and provides particularly effective training to enable a machine learning model to classify soil characteristics based on received hyperspectral reflectance spectra of the soil samples. The training method of Figure 7 thus begins, at step 702, by obtaining a hyperspectral reflectance profiles associated with a soil sample. This data is used as the input data in the subsequent training process. It will be appreciated that step 702 therefore corresponds to step 602 of Figure 6.

[0088] At step 704, the process obtains a target characteristic associated with the soil sample. This target characteristic may comprise soil-type classification information for the soil sample, such as a label or indicator of the primary constituent (e.g. a “clay” label). The target characteristic may alternatively or additionally comprise more detailed information concerning the proportions of one or more types of soil present in the soil sample, chemical composition information for the soil sample, and/or particle size information for the soil sample. The target characteristic is used as the target or ground truth in the training process. It will be appreciated that step 704 therefore corresponds to step 604 of Figure 6. Steps 706-712 then provide a more detailed description of the training process of step 606 of Figure 6.

[0089] Beginning at step 706, the hyperspectral reflectance profile obtained at step 702 is provided to a machine learning model. Responsive to this input, an output of the machine learning model is obtained at step 708. In particular, the output of the machine learning model comprises a determined characteristic of the soil sample, which is determined based on the input data. In other words, the machine learning model takes as its input a hyperspectral reflectance profile and produces as its output a classification of some characteristic of the soil sample. For example, the machine learning model may output a predicted primary constituent of the soil (e.g. “clay”), predicted constituent proportions (e.g. 10% clay, 90% sand), chemical property predictions and so on.

[0090] At step 710, an error between the determined characteristic of the soil sample and the target characteristic of the soil sample is determined. This can be performed in any suitable manner through comparison of the output obtained at step 708 and the target data provided at step 704. Responsive to this determination, the method proceeds at step 712 to adjust parameters ofthe machine learning model in order to reduce the error determined at step 710. [0091] The process of Figure Y can then repeated for further soil samples until a stopping criterion is met. For example, the method may be repeated N times, for a pre-set number of epochs or until a convergence criterion is satisfied.

Example implementation of Figure 7

[0092] To aid understanding, a specific implementation of the training process just described with reference to Figure 7 will now be provided. The method of Figure 7 was implemented by the present inventors as follows using a Jupyter Notebook and VS Code running PyTorch. The laptop had 32 GB of RAM with an Intel(R) Core(TM) i9-9980HK CPU @ 2.40GHz Computer Processing Unit (CPU). A dataset comprising 276 hyperspectral images of soil samples was obtained. Each image was 596 x 968 px. From these images, reflectance spectra for the set of soil samples were extracted and loaded into a dataloader . The spectra were split into training, validation and test data with a ratio of 60%, 20%, 20%. After randomly shuffling the data, approximately 15 Million training spectra, and 2.5 Million validation and test spectra were obtained. An Adaptive Moment Estimation (ADAM) optimiser with a weight decay of 10“⁵ and a cross entropy loss was used, with a batch size of 256 and a learning rate of 10“⁴. The network architecture comprised 299 inputs for each channel with four Fully Connected layers and Rectified Linear Unit (ReLU) activations to add non-linearity. The goal of the training was to train prediction of the primary constituent of the soil, which could either be clay, sand or silt. Hence, a final output layer of the network contained three outputs, one for each class, from which the maximum was taken as the “predicted class”. After training for only a few epochs, the method converged to very small losses and high accuracies in both training and validation. The results can be seen in Table 1 :

Table 1: Classification accuracy of each soil class in %

[0093] One issue with performing the method of Figure 7 for soil classification is that suitable hyperspectral images of soil samples that can be used to provide paired classifier training data (comprising hyperspectral reflectance spectra) are scarce. As noted, in the specific implementation described above only 276 hyperspectral images were available. This is significantly less than the thousands of images that are normally used during generalised learning of a neural network classifier. To address this issue, the present inventors have identified that results of the classifier training may be improved if a “pre-training” process is first used to pre-train the machine learning model (or a component thereof) via a reconstruction task in which it learns to “understand” hyperspectral images, and specifically hyperspectral reflectance spectra, better. Such a pre-training method is shown schematically in Figure 8.

[0094] The pre-training method of Figure 8 is effectively a reconstruction task, where the machine learning model is trained to generate (or “reconstruct”) target high resolution data from input low resolution data. In this particular example, because the aim is to train the model to “understand” hyperspectral reflectance spectra, the input and target data are hyperspectral reflectance spectra. The terms “low” and “high” should be understood as relative rather than absolute. In other words, these labels merely indicate that the high resolution data has a higher resolution than the low resolution data. For example, the high resolution hyperspectral reflectance spectra may have 66 channels while the low resolution hyperspectral reflectance spectra may have 33 or 22 channels. The below example will, for simplicity’s sake, assume that the entire machine learning model undergoes pre-training. However, as noted above, this is not essential and in some cases it may be preferable to pre-train only one or some components of the machine learning model, such as an encoder.

[0095] At step 802, the method begins by obtaining a target high resolution hyperspectral reflectance profile, such as through processing of a hyperspectral image. At step 804, a corresponding low resolution hyperspectral reflectance profile is obtained. This low resolution hyperspectral reflectance profile is typically obtained by down sampling or “binning” the high resolution hyperspectral reflectance profile obtained at step 802. For example, if the high resolution profile comprises 66 channels, the low resolution profile can be obtained by binning the profile down to a lower resolution, such as 33 or 22 channels.

[0096] At step 806, the low resolution hyperspectral reflectance profile is provided to the machine learning model as an input. Responsive to this input, an output of the machine learning model is obtained at step 808. In particular, the output of the machine learning model comprises a determined high resolution hyperspectral reflectance profile which the machine learning model predicts based on the input low resolution profile.

[0097] At step 810, an error between the determined high resolution hyperspectral reflectance profile and the target high resolution hyperspectral reflectance profile is determined. This can be performed in any suitable manner through comparison of the output obtained at step 808 and the target data provided at step 802. Responsive to this determination, the method proceeds at step 812 to adjust parameters of the machine learning model in order to reduce the error determined at step 810.

[0098] The process of Figure 8 can then be repeated for further hyperspectral reflectance spectra until a stopping criterion is met. For example, the method may be repeated N times, for a pre-set number of epochs or until a convergence criterion is satisfied.

[0099] A key benefit of the pre-training method of Figure 8 is that the target and input data obtained at steps 802 and 804 can be any suitable hyperspectral reflectance profile data, including unlabelled data. Critically, the present inventors have identified that this data does not need to relate to soil sample images to enable effective pre-training. Rather, hyperspectral reflectance spectra associated with any suitable hyperspectral images, such as hyperspectral satellite images, can be used. This data is widely available in large volumes, for example the PRISMA dataset contains hundreds of publicly available hyperspectral satellite images. This is in contrast to the classifier training which requires hyperspectral spectra of soil samples, which are in very short supply due to a lack of suitable hyperspectral images. Hence, the pre-training method of Figure 8 addresses the problem that labelled, hyperspectral soil spectra of the sort required for traditional classifier training are in short supply.

[00100] The present inventors have identified that pre-training a machine learning model using the method of Figure 8 is an effective way to prepare the machine learning model (or one or more components thereof) to subsequently perform the classification training of Figure 7 and ultimately the classification task of Figure 5. Accordingly, the steps of Figure 7 may follow the steps of Figure 8, and both Figures 7 and 8 may precede the steps of Figure 5. As noted above, in some examples the entire machine learning model pre-trained using the method of Figure 8 is subsequently used in the classifier training of Figure Y and the classification task of Figure 5. Alternatively, only a subset of components of the machine learning model is carried forward to the classifier training and eventual classification. Example implementation of Figure 8

[00101] To aid understanding, a specific implementation of the pre-training process just described with reference to Figure 8 will now be provided. The method of Figure 8 was implemented by the present inventors using an Implicit Neural Representation Network. In particular, a Sinus Implicit Neural Representation Network (SIREN) with Period Activation Functions was used to implement the pretraining. This network architecture is known in the art but has not been used in the context of soil classification.

[00102] The SIREN model was pretrained on unlabelled hyperspectral reflectance spectra obtained from hyperspectral images. In this example the hyperspectral images were obtained from the PRISMA dataset of hyperspectral satellite images. The PRISMA Satellite is a single satellite placed in suitable Lower Earth Orbit (LEO) and Sun-synchronous orbit (SSO) characterised by a repeat cycle of approximately 29 days. Its payload contains an imaging spectrometer (hyperspectral camera), capable of taking images in the VNIR and SWIR wavelength range from 400 to 2500 nm. The swath or Field of View (FOV) is 30km or 2.77 degrees. The VNIR range camera contains 66 bands from 400-101 Onm. The SWIR camera contains 173 bands from 920 - 2500 bands - a slight overlap in the NIR range from 920 - 101 Onm wavelengths. Hyperspectral reflectance spectra associated with each image obtained from this dataset were generated.

[00103] The obtained hyperspectral reflectance spectra were then binned (down sampled) from 66 channels to 33 channels to provide the required low resolution input data fortraining. The low resolution (33 channel) reflectance spectra were provided to the SIREN model and the model attempted to reconstruct the high resolution (66 channel) version. Through iteration and adjustment of the model parameters, the SIREN model was gradually trained to reconstruct high resolution hyperspectral reflectance spectra. Example inputs and outputs of this process are shown in Figures 9a-9c. Figure 9a shows a down sampled, low resolution (33 channel) reflectance profile provided as an input to the machine learning model during the pre-training. Figure 9b shows the output of the machine learning model, which is a determined (predicted) high resolution version of the reflectance profile. Figure 9c represents the ground truth, i.e. the original real high resolution reflectance profile prior to down sampling. Through training, the machine learning model iteratively improves at reconstructing the ground truth high resolution profile from the low resolution input profile.

[00104] In the present example implementation, the training set contained ~ 10 million pixels and the corresponding spectra in the VNIR range. The hyperparameters used were a learning rate of 10“⁵ , weight decay of 10“⁴ , with an ADAM optimiser. Finally, both the Evidence Lower Bound (ELBO) loss as well as a combination of Mean-Squared-Error (MSE) and L1 loss were tested. The batch sizes were 32 for training and validation and 16 for the test set. The network architecture was then adapted with convolutional layers, slowly building up the number of parameters and layers, resulting in a CNN with residual layers.

[00105] The network architecture used during pre-training in this example can be described in two parts, as shown in Figure 10a. The network consists of an encoder 1002 with convolutional layers, and a decoder 1004, hence the network was analogous to an auto-encoder. Pre-training using this network architecture continued for 25 Epochs using the steps described above in relation to Figure 8.

Use of pre-trained encoder to perform classification training

[00106] As noted above, one or more components pre-trained using the method of Figure 8 can subsequently be used to implement the classification training of Figure 7. This process was implemented by the present inventors, as follows.

[00107] Once pre-training was completed in the manner just described, the pre-trained encoder 1002 was then implemented in a classification architecture, shown in Figure Kb. Now the encoder 1002 was paired with a classification head K06 consisting of Fully Connected (FC) layers and having three outputs (again corresponding to the soil labels “clay”, “silt” and “sand” in this case). This architecture was then used to implement a classification training regime as described above in relation to Figure Y, to train the model to classify hyperspectral reflectance spectra into one of three soil categories - clay, silt and sand. [00108] The same hyperparameters as were used in the pre-training were maintained: ADAM optimiser with a learning rate of 10“⁵ and weight decay of 10“⁴ . Batch sizes of 32 for training and validation set and 16 for the test set were used. The samples were split into training, validation and test set proportions 60% - 20% - 20%. In the present example, the weights in the pre-trained encoder were initially frozen during classifier training, allowing the first layer and classification head to train a few iterations before unfreezing the weights and allowing the backward pass to change all weights end-to-end throughout the network. The training continued for 20 Epochs.

[00109] The present inventors identified that the pre-training process exemplified by Figure 8 improved the ability of the machine learning model to subsequently classify hyperspectral reflectance spectra in respect of soil characteristics. Specifically, the results showed great promise in using reconstruction as an unsupervised pretraining methodology prior to soil classification.

[00110] The present inventors also identified that a particularly effective network architecture for use during the classifier training comprises of a convolutional neural network (CNN) with residual layers, also known as a ResNet, where the encoder has preferably been pre-trained according to the method of Figure 8 outlined above.

[00111] With reference to Figure 11 , a computing device 1100 suitable for carrying out the methods described above will now be described. Figure 11 shows a block diagram of one implementation of a processing system 1100 in the form of a computing device within which a set of instructions for causing the computing device to perform any one or more of the methodologies discussed herein, may be executed. In alternative implementations, the computing device may be connected (e.g., networked) to other machines in a Local Area Network (LAN), an intranet, an extranet, or the Internet. The computing device may operate in the capacity of a server or a client machine in a client-server network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. The computing device may be a personal computer (PC), a tablet computer, a set-top box (STB), a Personal Digital Assistant (PDA), a cellular telephone, a web appliance, a server, a network router, switch or bridge, or any machine capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by that machine. Further, while only a single computing device is illustrated, the term “computing device” shall also be taken to include any collection of machines (e.g., computers) that individually or jointly execute a set (or multiple sets) of instructions to perform any one or more of the methodologies discussed herein.

[00112] The example processing system 1100 includes a processor 1102, a main memory 1104 (e.g., read-only memory (ROM), flash memory, dynamic random access memory (DRAM) such as synchronous DRAM (SDRAM) or Rambus DRAM (RDRAM), etc.), a static memory 1106 (e.g., flash memory, static random access memory (SRAM), etc.), and a secondary memory (e.g., a data storage device 1118), which communicate with each other via a bus 1130.

[00113] Processor 1102 represents one or more general-purpose processors such as a microprocessor, central processing unit, or the like. More particularly, the processor 1102 may be a complex instruction set computing (CISC) microprocessor, reduced instruction set computing (RISC) microprocessor, very long instruction word (VLIW) microprocessor, processor implementing other instruction sets, or processors implementing a combination of instruction sets. Processor 1102 may also be one or more special-purpose processors such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), network processor, or the like. Processor 1002 is configured to execute the processing logic (instructions 1122) for performing the operations and steps discussed herein.

[00114] The processing system 1100 may further include a network interface device 1008. The processing system 1100 also may include a video display unit 1110 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)), an alphanumeric input device 1112 (e.g., a keyboard or touchscreen), a cursor control device 1114 (e.g., a mouse or touchscreen), and an audio device 1016 (e.g., a speaker).

[00115] It will be apparent that some features of the processing system 1100 shown in Figure 10 may be absent. For example, the processing system 1100 may have no need for display device 1110 (or any associated adapters). This may be the case, for example, for particular server-side computer apparatuses which are used only for their processing capabilities and do not need to display information to users. Similarly, user input device 1112 may not be required. In its simplest form, processing system 1100 comprises processor 1102 and main memory 1104.

[00116] The data storage device 1118 may include one or more machine-readable storage media (or more specifically one or more non-transitory computer-readable storage media) 1128 on which is stored one or more sets of instructions 1122 embodying any one or more of the methodologies or functions described herein. The instructions 1122 may also reside, completely or at least partially, within the main memory 1104 and/or within the processor 1002 during execution thereof by the processing system 1100, the main memory 1104 and the processor 1102 also constituting computer-readable storage media 1128.

[00117] The various methods described above may be implemented by a computer program. The computer program may include computer code arranged to instruct a computer to perform the functions of one or more of the various methods described above. The computer program and/or the code for performing such methods may be provided to an apparatus, such as a computer, on one or more computer readable media or, more generally, a computer program product. The computer readable media may be transitory or non-transitory. The one or more computer readable media could be, for example, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, or a propagation medium for data transmission, for example for downloading the code over the Internet. Alternatively, the one or more computer readable media could take the form of one or more physical computer readable media such as semiconductor or solid state memory, magnetic tape, a removable computer diskette, a random access memory (RAM), a read-only memory (ROM), a rigid magnetic disc, and an optical disk, such as a CD-ROM, CD-R/W or DVD.

[00118] The computer program is executable by the processor 1102 to perform functions of the systems and methods described herein.

[00119] In an implementation, the modules, components, and other features described herein can be implemented as discrete components or integrated in the functionality of hardware components such as ASICS, FPGAs, DSPs, or similar devices.

[00120] A “hardware component” is a tangible (e.g., non-transitory) physical component (e.g., a set of one or more processors) capable of performing certain operations and may be configured or arranged in a certain physical manner. A hardware component may include dedicated circuitry or logic that is permanently configured to perform certain operations. A hardware component may be or include a specialpurpose processor, such as a field programmable gate array (FPGA) or an ASIC. A hardware component may also include programmable logic or circuitry that is temporarily configured by software to perform certain operations.

[00121] Accordingly, the phrase “hardware component” should be understood to encompass a tangible entity that may be physically constructed, permanently configured (e.g., hardwired), or temporarily configured (e.g., programmed) to operate in a certain manner or to perform certain operations described herein.

[00122] In addition, the modules and components can be implemented as firmware or functional circuitry within hardware devices. Further, the modules and components can be implemented in any combination of hardware devices and software components, or only in software (e.g., code stored or otherwise embodied in a machine-readable medium or in a transmission medium).

[00123] Unless specifically stated otherwise, as apparent from the following discussion, it is appreciated that throughout the description, discussions utilizing terms such as "receiving”, “determining”, “comparing”, “enabling”, “maintaining,” “identifying,”, “receiving”, “providing” or the like, refer to the actions and processes of a computer system, or similar electronic computing device, that manipulates and transforms data represented as physical (electronic) quantities within the computer system's registers and memories into other data similarly represented as physical quantities within the computer system memories or registers or other such information storage, transmission or display devices.

[00124] In the examples provided above, the systems and methods are configured to determine soil classification information, such as soil type. However, it will be appreciated that the systems and methods described herein may, additionally or alternatively, be configured to provide information regarding particle size and/or chemical composition of a soil sample. For example, in addition to determining a soil type, e.g. clay, silt, or sand, embodiments of the present disclosure may also be configured to determine a secondary descriptor, such as particle size. Moreover, in addition to determining soil type based on a hyperspectral reflectance profile, systems and methods of the present disclosure may be configured to identify or estimate one or more chemical components of the sample.

[00125] It is to be understood that the above description is intended to be illustrative, and not restrictive. Many other implementations will be apparent to those of skill in the art upon reading and understanding the above description. Although the present disclosure has been described with reference to specific example implementations, it will be recognized that the disclosure is not limited to the implementations described but can be practiced with modification and alteration within the scope of the appended claims.

Accordingly, the specification and drawings are to be regarded in an illustrative sense rather than a restrictive sense. The scope of the disclosure should, therefore, be determined with reference to the appended claims.

[00126] While at least one exemplary embodiment has been presented in the foregoing detailed description, it should be appreciated that a vast number of variations exist. It should also be appreciated that the exemplary embodiment or exemplary embodiments are only examples and are not intended to limit the disclosure in any way.

Claims

1 . A method for predicting soil characteristic information, the method comprising: receiving data representative of a hyperspectral image of a soil sample; determining a hyperspectral reflectance profile for the soil sample based on the received data, wherein the hyperspectral reflectance profile comprises reflectance intensity information; determining one or more soil characteristics based on the determined hyperspectral reflectance profile of the soil sample, wherein the one or more soil characteristics include one or more of: soil-type classification information for the soil sample; chemical composition information for the soil sample; particle size information for the soil sample.

2. The method of claim 1 , wherein the determining one or more soil characteristics based on the determined hyperspectral reflectance profile of the soil sample comprises comparing the determined hyperspectral reflectance profile to one or more known reflectance spectra.

3. The method of claim 1 or 2, wherein the hyperspectral reflectance profile comprises reflectance intensity information for wavelengths in the range of:

400nm to 2000nm;

400nm to 1000nm, or between 400nm to 550nm and between 850nm to 1000nm.

4. The method of claim 1 , 2 or 3, wherein the hyperspectral image of a soil sample is a snap-shot hyperspectral image.

5. The method of any of claims 1 to 4, wherein the hyperspectral image comprises a spectral sampling interval of approximately 20nm or less, approximately 15nm or less, approximately 10nm or less, or approximately 5nm or less, and wherein the sampling interval optionally extends across one or more of the ranges of claim 2

6. The method according to any of claims 1 to 5, wherein the method further comprises: identifying one or more regions of interest, ROIs, in the hyperspectral image; determining the hyperspectral reflectance profile in the one or more ROIs.

7. The method according to claim 6, wherein a plurality of ROIs is identified, and wherein the determining the hyperspectral reflectance profile of the ROIs comprises determining a mean hyperspectral reflectance profile for the plurality of ROIs.

8. The method according to any of claims 1 to 7, wherein the hyperspectral image is captured at a distance D from the soil sample, and wherein D is approximately 100m or less, optionally approximately 25m or less, optionally approximately 10m or less, optionally approximately 5m or less, optionally approximately 1 m or less.

9. The method according to any of claims 1 to 8, wherein the method further comprises the step of capturing the hyperspectral image.

10. The method according to claim 9, wherein the image is captured from a soil sample in situ or wherein the image is captured from a soil sample in a test environment.

11 . The method according to any of claims 1 to 10, wherein the soil-type classification information comprises a soil classification label, such as clay, silt, and/or sand.

12. The method according to any of claims 1 to 11 , wherein the step of determining one or more soil characteristics based on the determined hyperspectral reflectance profile of the soil sample comprises using a machine learning model to predict the soil characteristics.

13. The method according to claim 12, further comprising the steps of: training the machine learning model to identify a characteristic of a soil sample according to any of claims 14 to 22.

14. A method of training a machine learning model to identify a characteristic of a soil sample, the method comprising: obtaining a hyperspectral reflectance profile associated with a soil sample; obtaining a target characteristic associated with the soil sample; providing the hyperspectral reflectance profile to a machine learning model to obtain an output of the machine learning model, the output of the machine learning model comprising a determined characteristic of the soil sample; and adjusting parameters of the machine learning model to reduce an error between the determined characteristic of the soil sample and the target characteristic associated with the soil sample.

15. The method of claim 14, wherein the hyperspectral reflectance profile comprises reflectance data for wavelengths in the range of:

400nm to 2000nm;

400nm to 10OOnm, and between 400nm to 550nm and between 850nm to 1000nm.

16. The method of claim 14 or 15, wherein the determined and target characteristics of the soil sample comprise one or more of: soil-type classification information for the soil sample; chemical composition information for the soil sample; particle size information for the soil sample.

17. The method of any of claims 14 to 16, further comprising, prior to providing the hyperspectral reflectance profile to the machine learning model, performing a pre-training process on one or more components of the machine learning model, the pre-training process comprising: obtaining a target high resolution hyperspectral reflectance profile; obtaining a low resolution hyperspectral reflectance profile; providing the low resolution hyperspectral reflectance profile to the one or more components of the machine learning model to obtain an output of the one or more components of the machine learning model, the output comprising a determined high resolution hyperspectral reflectance profile; and adjusting parameters of the one or more components of the machine learning model to reduce an error between the determined high resolution hyperspectral reflectance profile and the target high resolution hyperspectral reflectance profile.

18. The method of any of claims 14 to 17, wherein the machine learning model comprises an encoder.

19. The method of any of claims 14 to 18, wherein the machine learning model comprises an implicit neural representation network.

20. The method of claim 19, wherein the machine learning model comprises a sinus implicit neural representation network.

21 . The method of any of claims 14 to 20, wherein the machine learning model comprises a residual network.

22. The method of any of claims 14 to 21 , wherein the machine learning model is further trained with hyperspectral image data comprising one more labels indicative of sample wetness.

23. A method of obtaining an estimate of soil classification using a machine learning model trained in accordance with any of claims 14 to 22, the method comprising: obtaining input data comprising data representative of a hyperspectral image of a soil sample; applying the input data to the machine learning model to obtain an output of the machine learning as the estimate.

24. A system comprising: one or more processors; one or more memories having stored thereon computer readable instructions configured to cause the one or more processors to perform operations comprising the steps of any of the preceding claims.

25. The system of claim 24, further comprising one or more sensors, wherein the one or more sensors optionally comprises a hyperspectral camera.

26. One or more computer readable media comprising instructions, that, when executed by one or more data processing apparatus, cause the one or more data processing apparatus to perform operations comprising the steps of any of claims 1 to 23.

27. A machine learning model stored on one or more computer readable media, wherein the model has been trained according to the method of claims 14 to 22 or the method of claim 13.

28. A system for capturing hyperspectral reflectance information of a soil sample, the system comprising: a remotely operated vehicle, ROV, comprising one or more hyperspectral cameras configured to capture a hyperspectral image of a soil sample; wherein the one or more hyperspectral cameras is configured to detect a hyperspectral reflectance profile comprising reflectance intensity information for wavelengths in the range of:

400nm to 2000nm;

400nm to 1000nm, or between 400nm to 550nm and between 850nm to 10OOnm

29. The system of claim 28, wherein the ROV comprises: a light source, and, optionally a calibration surface.

30. The system of claim 28 or 29, further comprising: a communication module configured to transmit hyperspectral image data to a remote storage module.

31 . The system of any of claims 28 to 30, wherein the system further comprises one or more processors configured to carry out the steps of any of claims 1 to 13 or 23.