CN115984702B - Method for monitoring forest land change of mining area based on high-resolution images - Google Patents
Method for monitoring forest land change of mining area based on high-resolution images Download PDFInfo
- Publication number
- CN115984702B CN115984702B CN202310082173.1A CN202310082173A CN115984702B CN 115984702 B CN115984702 B CN 115984702B CN 202310082173 A CN202310082173 A CN 202310082173A CN 115984702 B CN115984702 B CN 115984702B
- Authority
- CN
- China
- Prior art keywords
- image
- resolution
- change detection
- forest land
- woodland
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a method for monitoring forest land change in a mining area based on high-resolution images, which comprises the following steps: acquiring an existing forest land change detection data set and a high-resolution seventh satellite image data set; preprocessing a high-resolution seventh satellite image data set to obtain a high-resolution woodland change detection training set and a high-resolution woodland change detection test set; the existing forest land change detection data set is respectively subjected to image alignment with the high-resolution forest land change detection training set and the high-resolution forest land change detection test set based on the image alignment module; training the deep learning model based on the aligned existing forest land change detection data set, adjusting the model according to the aligned high-resolution forest land change detection training set, and carrying out change detection on the aligned high-resolution forest land change detection test set according to the adjusted model to obtain mining area forest land change distribution data. The method solves the problem of suitability of the training set image and the existing image, and rapidly and accurately monitors the forest land change information of the mining area on the high-resolution seventh image.
Description
Technical Field
The invention relates to the technical field of change detection, in particular to a method for monitoring forest land change in a mining area based on high-resolution images.
Background
Forest land protection is not only a main way for increasing carbon sink, but also a necessary condition for realizing a double-carbon target in China early days. While woodland destruction is more the second largest source of carbon dioxide emissions in the world today next to fossil fuel combustion. In addition, the forest land is also an important factor influencing the ecological environment, and plays an important role in climate control, water and soil conservation, biodiversity protection, wind prevention, sand fixation and the like. The forest land protection is practically realized, and the realization of double-carbon target and ecological civilization construction can be effectively promoted. Under the influence of mining development, forest lands on the surface of a mining area are easy to damage, and the reduction of the coverage area, the reduction of the quantity and the like of the forest lands are mainly shown. Therefore, monitoring the change situation of the forest land in the mining area is an important link for realizing the forest land protection.
At present, the monitoring of the forest land change of the mining area is carried out by means of manual step-by-step filling statistics, unmanned aerial vehicle spot check, online video monitoring and the like. Wherein, the manual step-by-step filling mode has the possibility of certain missing report, false report and hidden report, and the data quality is difficult to be ensured; the unmanned aerial vehicle has high flight cost, can only carry out sampling inspection, and cannot be comprehensively monitored. On-line video is usually only installed in local areas, cannot cover the full view, and is costly to invest and maintain. The condition of the sparse areas of the forest land such as mining areas cannot be clearly reflected by the condition of the sparse areas of the forest land such as mining areas, and the condition of the satellite data with high resolution is mainly identified by manually interpreting and identifying the satellite data with high resolution, namely, the satellite data with high resolution and the satellite data with high resolution are greatly influenced by personnel interpretation experience. With successful transmission and use of higher resolution sub-m-level high-resolution satellite No. seven, it is necessary to develop a mining area forest land change monitoring method based on data of the satellite No. seven so as to improve efficiency, accuracy and economy of mining area forest land change monitoring.
In the middle-low resolution image era, the traditional machine learning method is widely applied to forest land change detection. The method can integrate various knowledge and rules, has strong mechanization, is very sensitive to experience knowledge, characteristic parameters and the like, and has poor universality and mobility. In addition, due to insufficient consideration of demand information, detection results of the method on high-resolution images often form spiced salt shapes. Along with the development of deep learning, a great number of forest land change detection methods based on deep learning are proposed, and the methods often adopt a twin network model, and the model is trained by a great number of image pairs and corresponding forest land change labels, so that the model can autonomously adjust parameters to simulate the nonlinear relationship between the image pairs and the labels, thereby obtaining a result. However, such methods often require a large number of training samples, and the difficulty in acquiring data in practical application is high. The existing forest land change detection data set is mostly manufactured based on middle-low resolution images such as Landsat and the like, and no forest land change detection data set based on high-resolution images is available. Meanwhile, as the features of the ground features on the remote sensing images of different regions have significant differences, the model trained on a certain region image cannot well act on other region images with significant differences, namely, how to accurately realize the forest land change detection of the high-resolution seventh image by using the universal model under the condition of only having a small amount of high-resolution samples becomes the key point of the solution.
Disclosure of Invention
In view of the above, the invention provides a method for monitoring the change of the mine forest land based on the high-resolution images, which aims at solving the problem that a large amount of sample data with medium and low resolution cannot be directly applied to the identification of the high-resolution seventh image.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
a method for monitoring changes in mine timbering based on high-resolution images, comprising:
acquiring an existing forest land change detection data set and a high-resolution seventh satellite image data set;
preprocessing a high-resolution seventh satellite image data set to obtain a high-resolution woodland change detection training set and a high-resolution woodland change detection test set;
the method comprises the steps that based on an image alignment module, an existing woodland change detection data set is respectively subjected to image alignment with a high-order woodland change detection training set and a high-order woodland change detection test set, so that an aligned existing woodland change detection data set, an aligned high-order woodland change detection training set and an aligned high-order woodland change detection test set are obtained;
training the deep learning model based on the aligned existing forest land change detection data set, adjusting the deep learning model according to the aligned high-resolution forest land change detection training set, and carrying out change detection according to the aligned high-resolution forest land change detection test set by the adjusted deep learning model to obtain mining area forest land change distribution data.
Preferably, preprocessing the high score seventh satellite image dataset includes:
carrying out radiometric calibration and geometric correction, single-view image fusion, geometric fine correction, image mosaic, image registration, image clipping, image pair generation, note making and data division on the high-resolution seventh satellite image data.
Preferably, the image alignment module comprises an image converter and an image discriminator, wherein the image converter comprises 3 2-dimensional convolution layers, and the image discriminator comprises 2-dimensional convolution layers and 2 fully-connected layers;
preferably, the objective function of the image alignment module is:
wherein T represents an image converter, D represents an image discriminator, U represents a change detection dataset from a woodland, V represents a change detection training set for a high-resolution woodland or a change detection test set for a high-resolution woodland, D (T (x)) represents a discrimination result of the image discriminator on the image T (x) after the conversion of the image converter, P represents a probability, a first term after the equation represents a probability that the image discriminator discriminates an image converted from an existing dataset into an image of the existing dataset, and a second term represents a probability that the image discriminator discriminates the image converted from the high-resolution dataset into an image of the high-resolution dataset.
Preferably, the method further comprises the step of carrying out data standardization processing on the existing forest land change detection data set, the high-branch forest land change detection training set and the high-branch forest land change detection test set.
Preferably, the deep learning model is a twin U-Net model and comprises an encoder and a decoder, wherein the encoder comprises 5 submodules, each submodule comprises 2 convolution layers with the convolution kernel size of 3×3, and downsampling is carried out through 1 max pooling layer after the convolution layers of the first 4 submodules;
the decoder comprises 5 sub-modules, each sub-module comprises 2 convolution layers with convolution kernel size of 1×1, and the back 4 sub-modules each perform up-sampling through 1 pooling layer after convolution layers.
Preferably, training the deep learning model based on the aligned existing forest land change detection data set specifically includes:
carrying out convolution and downsampling operations on the aligned pre-woodland time-phase image and the aligned post-woodland time-phase image for 2 times through a first layer submodule of an encoder respectively, and fusing the pre-time-phase image features and the post-time-phase image features extracted by the first layer;
respectively carrying out convolution and downsampling operations for 2 times on the pre-time-phase image features and the post-time-phase image features extracted by the first layer submodule through the second layer submodule, and fusing the pre-time-phase image features and the post-time-phase image features extracted by the second layer submodule; and so on, until proceeding to the fifth layer sub-module;
inputting the fusion result of the front time phase image feature and the rear time phase image feature output by the fifth layer sub-module of the encoder into the fifth layer sub-module of the decoder, and inputting the fusion result into the fourth layer sub-module of the decoder after convolution operation and up-sampling;
and the fourth layer sub-module of the decoder combines the output result of the fifth layer with the fusion result of the front time phase image feature and the rear time phase image feature extracted by the same layer of the encoder, and inputs the result to the third layer sub-module of the decoder after convolution operation and up-sampling, and the like until the result reaches the first layer sub-module, and outputs a woodland change detection prediction result. .
The invention has the following advantages:
1. the method realizes the alignment technology of the high-resolution seventh image through conversion and discrimination, further obtains the improved change label, can be applied to the rapid detection of the high-resolution data on the woodland, and has higher accuracy.
2. The method can support and realize effective supervision of the mine forest land, protect and effectively restore the ecological environment of the mine, and has a certain significance for maintaining the ecological safety of vegetation of the mine and guaranteeing the green high-quality stable development of the mine.
3. Compared with modes of manual investigation and reporting, video monitoring, unmanned aerial vehicle monitoring and the like, the technical method is simple to operate, low in cost and high in accuracy, can obviously reduce the manual monitoring cost and improves the monitoring frequency.
4. The method can provide reference for low-resolution to high-resolution monitoring of other ground object types, and has important significance for improving accuracy and developing rapid remote sensing monitoring technology method application.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only embodiments of the present invention, and that other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a method for monitoring forest land changes in mining areas based on high-resolution images.
Fig. 2 is a schematic diagram of remote sensing images in different data sets according to the present invention, in which fig. 2 (a) is a conventional forest land change detection data set, and fig. 2 (b) is a high-resolution seventh satellite image.
Fig. 3 is a schematic diagram of an image alignment module according to the present invention.
FIG. 4 is a schematic diagram of a twin U-Net model provided by the invention.
FIG. 5 is a graph showing the results of the forest land change detection on the test set sample provided by the present invention, wherein FIG. 5 (a) is a 2021 month 6 image, FIG. 5 (b) is a 2022 month 1 image, FIG. 5 (c) is a forest land change detection label, FIG. 5 (d) is the results of the present invention, and FIG. 5 (e) is the results of the model training using only the high-score training set;
FIG. 6 is a graph showing a comparison of model parameters before and after image alignment according to the present invention.
FIG. 7 is a graph showing a comparison of model parameters performed by the conventional method provided by the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The embodiment of the invention discloses a method for monitoring forest land changes in a mining area based on high-resolution images, which is shown in fig. 1 and comprises the following steps:
acquiring an existing forest land change detection data set and a high-resolution seventh satellite image data set;
preprocessing a high-resolution seventh satellite image data set to obtain a high-resolution woodland change detection training set and a high-resolution woodland change detection test set;
the method comprises the steps that based on an image alignment module, an existing woodland change detection data set is respectively subjected to image alignment with a high-order woodland change detection training set and a high-order woodland change detection test set, so that an aligned existing woodland change detection data set, an aligned high-order woodland change detection training set and an aligned high-order woodland change detection test set are obtained;
training the deep learning model based on the aligned existing forest land change detection data set, adjusting the deep learning model according to the aligned high-resolution forest land change detection training set, and carrying out change detection according to the aligned high-resolution forest land change detection test set by the adjusted deep learning model to obtain mining area forest land change distribution data.
In this embodiment, as shown in fig. 2 (a), to obtain the existing forest land change detection dataset, fig. 2 (b) is a high-resolution seventh image of different time phases covering the area of the mining area. The existing forest land change detection data set can be a PRODES (Brazilian Amazon rain forest satellite monitoring program) data set and the like, the method is mainly used for assisting the training of the deep learning model, and the requirements on the deep learning model mainly comprise the following two aspects: 1. there are enough samples, each sample includes the image pairs of the front, back time phase and its corresponding forest land change label; 2. the spatial resolution of the image in the sample cannot exceed 30 meters. The high-score No. 7 satellite image is mainly used for fine tuning and performance evaluation of a model, the acquired image data is a high-score No. 7 image L1A product, and the following requirements should be paid attention to during acquisition: 1. the imaging time difference of the images in the same time phase range can not exceed 15 days according to the task selection of different time phase ranges. 2. The images within any one phase range should cover the mine completely.
In this embodiment, the acquired satellite image with No. seven high score is processed to form the change detection data set of the forest land with high score, and then the change detection data set can be applied to fine tuning and performance evaluation of the model. The treatment process mainly comprises the following steps:
(1) Radiometric calibration and geometric correction: and carrying out radiometric calibration and geometric correction on all acquired high-resolution seventh satellite image data according to the self-contained calibration coefficient and RPC file, wherein the technology does not involve any quantitative inversion work, so that atmospheric correction is not required.
(2) And (3) single-view image fusion: for all high-resolution seventh images, the GS (Gram-Schmidt) tool in ENVI software was used to fuse the visible and full-color band images to obtain true color images with a resolution of 0.8 meters.
(3) Geometric fine correction: in order to ensure the spatial accuracy of the detection result, geometric fine correction needs to be performed on all high-resolution images by using a ground control point, and an Image to map tool of ENVI software is adopted in the geometric fine correction, so that the root mean square error is required to be smaller than 1 pixel.
(4) And (3) image mosaic: an ENVI Seamless Mosaic tool (Seamless Mosaic) is used to Mosaic and uniformly color images of front and back time phases.
(5) Image registration: an Image automatic registration tool Image to Image in ENVI is used for registering the post-mosaic Image of the rear time phase to the post-mosaic Image of the front time phase, the registration error is smaller than 1 pixel, and when the error is larger, the control points can be manually selected for registration.
(6) Cutting the images: and cutting the mosaic images of the front time phase and the rear time phase respectively by using a vector file representing the mining area range to obtain high-resolution No. 7 images of the front time phase and the rear time phase of the research area range.
(7) And (3) generating an image pair: the high-resolution seventh image of the front and rear time phases is cut into a plurality of 256×256 pixel tiles, respectively, and then the front and rear time phase tiles corresponding to the same region are combined together to form an image pair.
(8) And (3) label manufacturing: about 300 image pairs are selected for which to construct a woodland variation label. The labels and the images in the image pairs have the same size, the corresponding area of the image pair, where the forest land changes, is marked as 1, the other areas are marked as 0, and tools such as LabelMe or ArcGIS can be adopted in manufacturing.
(9) Dividing data: the single image pair and the label thereof form a sample, 50% of the sample is divided into a high-resolution woodland change detection training set for fine adjustment of the model, and 50% of the sample is divided into a high-resolution woodland change detection test set for evaluation of the performance of the model.
In this embodiment, in order to accelerate the convergence rate of the model, the deep learning method often requires that the input data is normalized. Therefore, after preprocessing the high-resolution seventh satellite image data set, the existing woodland change detection data set, the high-resolution woodland change detection training set and the high-resolution woodland change detection test set are subjected to standardized processing, namely, the average value of the image pixels is subtracted from all pixels on a single image, and then the average value is divided by the standard deviation of the image pixels.
In this embodiment, the image alignment module is used to align the image of the existing forest land change detection dataset and the image of the high-resolution seventh satellite dataset after the standardized processing, as shown in fig. 3:
the image alignment module converts the high-resolution satellite images of the corresponding time phases and the existing dataset images into the same data distribution, so that the images of the front time phase and the rear time phase need to be processed separately. The module adopts the idea of antagonizing the generation network and comprises two sub-modules of an image converter and an image discriminator. The image converter comprises 3 2-dimensional convolution layers (the convolution layers have the size of 3×3 and the number of the convolution layers is 32), and is responsible for converting the existing woodland change detection data set image and the high-resolution seventh satellite data set image of the same time phase into the same data distribution. The image discriminator consists of 2-dimensional convolution layers (convolution kernel size 3×3, number 32) and two full-connection layers (neuron numbers 100 and 1, respectively). Which is used to determine from which data set the converted image is derived (high score satellite data set number seven or existing forest land change detection data set). The two sub-modules train step by step and game each other. The parameters of the arbiter are fixed while training the converter, with the aim of making it impossible for the image arbiter to determine from which data set the converted image is derived. When training the image discriminator, the image converter is fixed, and the training goal is to make the image discriminator identify to which data set the converted image belongs as much as possible. Thus, its objective function can be described as follows:
wherein T represents an image converter, D represents an image discriminator, U represents a data set from the prior art, V represents a high-resolution woodland change detection training set or a high-resolution woodland change detection test set. P represents probability. The first term after the equation represents the probability that the discriminator discriminates the image from the existing dataset converted by the converter as an image of the existing dataset, and the second term represents the probability that the discriminator discriminates the image from the high-score dataset converted by the converter as an image of the high-score dataset. The decision is trained such that the sum of the two probabilities is maximized and the converter is trained such that the sum of the two probabilities is minimized.
The two are separately and alternately trained, and the training of the converter is started, then the discriminant is trained, and the converter and the discriminant are trained again, and the two are sequentially and alternately performed until the prescribed turn. With the increase of training rounds, the capacities of the discriminator and the converter are mutually promoted in the game, and finally the discriminator can better distinguish different data distributions, and the converter can also convert images of different data distributions into the same data distribution. The objective function of this training is set as follows:
F＝xlog(p)+(1-x)log(1-p)
where x is the home label of the image (the existing dataset image is 1, the high-score seventh image is 0), and p is the probability that the converted image comes from the existing dataset. Essentially, F is a binary cross entropy function that indicates whether the transformed image belongs to correctly determined. (the larger the F value, the more accurately the judgment is made, the smaller the F value, the more the judgment is made.
Because the aligned high-resolution seventh image and the aligned existing data set have the same or similar image data distribution, a small amount of aligned high-resolution image samples can be used for easily transferring the pre-trained model into the high-resolution seventh image for forest land change detection, as shown in fig. 6; the forest land change detection is carried out by other conventional methods, the twin network model is directly pre-trained on a certain existing data set, and then a small amount of high-score samples are used for migration learning to obtain the following steps: the large difference between the data distribution of the data set constructed based on the low-resolution images such as Landsat and the high-resolution seventh image results in a poor model migration learning effect on the existing data set by using a small number of high-resolution seventh image samples, as shown in FIG. 7, wherein the existing data set images specifically refer to the high-resolution forest land change detection training set.
In this embodiment, the aligned existing forest land change detection data set is used to train the deep learning model:
as shown in fig. 4, the deep learning model is a twin U-Net model, including an encoder and a decoder. The encoder is mainly used for extracting layer characteristics of images in front and back time phases. It contains 5 sub-modules, each comprising 2 convolution layers of convolution kernel size 3 x 3. Each sub-module is followed by a max pooling layer to downsample the features extracted by the sub-modules except the last sub-module to reduce the size of the feature map. The front and back time phase images share one encoder, and the features of the front and back time phase images extracted by each sub-module are fused to be input into the decoder for use. The decoder and the encoder form a symmetrical structure and are also composed of 5 submodules with the same structure. Except that each sub-module is followed by an up-sampling layer to enlarge the size of the feature map, except for the uppermost sub-module. The input of each sub-module in the decoder is composed of two parts, one part is the output of the lower sub-module amplified by the up-sampling layer, and the other part is the fusion output characteristic of the sub-module of the corresponding level in the encoder. The decoder comprehensively uses the characteristics of different levels to generate a woodland change detection prediction result with the same size as the original image. Training the model by using all data in the aligned woodland change detection data set, wherein an Adam optimizer with an initial learning rate of 0.001 is adopted, an objective function is set to be binary cross entropy, and the training times are 100 times.
Firstly, respectively extracting characteristic information in front and back time phase images by using a continuous convolution pooling layer, gradually mapping the characteristic information to high dimension, and fusing the extracted front and back time phase image characteristics of each layer; and then mapping the high-dimensional characteristics to the low dimension again by a convolution method, fusing the images which are identical to the dimensions of the high-dimensional characteristics in the shrinkage network under the same dimension, wherein the dimensions can be 2 times of the original dimensions in the fusion process, and then, multiple convolutions are needed, so that the dimensions after processing are identical to the dimensions before the fusion operation, the processed dimensions can be fused with the images under the same dimension for the second time after the convolution operation is performed, and finally, the forest land change detection result which is identical to the dimensions of the original images is output.
In this embodiment, after passing through the image alignment module, the high-resolution image and the existing dataset image can be more consistent in data distribution. However, in order to make the trained model in the previous step better perform forest land change area detection on the high-resolution image, fine tuning needs to be performed on the trained model, that is, the trained model is trained again through the aligned high-resolution forest land change detection training set. The same setting as the previous step is adopted during training.
In this embodiment, the detection result is subjected to cross-correlation index evaluation, after the related requirements are met, the existing forest land dataset covering the mining area and the high-resolution seventh image are subjected to image alignment according to the steps, and then input into a training and adjusting model to obtain the mining area forest land change result based on two-stage high-resolution seventh image monitoring.
The effectiveness of the technique was verified by using a high score No. 7 image covering the a mine. Wherein mining area a uses high-score image data No. 7 of 2021, 6 and 2022, 1, as shown in fig. 5 (a) and 5 (b), a total of 300 samples of 256×256 were constructed manually, 150 samples being used as a training set for fine-tuning the model, and 150 samples being used as a test set for evaluating the model performance. The auxiliary data adopts a data set of sentinel No. 2 from 2016 to 2017 of PRODES (Amazon rain forest satellite monitoring program of Brazil), and 1400 256×256 samples are screened out. The results show that the results of the forest land change detection (shown as fig. 5 (d), the cross-over ratio of 93%) by the present invention are better than the results of model training by using a small number of high-score training sets alone (shown as fig. 5 (e), the cross-over ratio of 81%).
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (6)
1. A method for monitoring changes in mine timberland based on high-resolution images, comprising:
acquiring an existing forest land change detection data set and a high-resolution seventh satellite image data set;
preprocessing a high-resolution seventh satellite image data set to obtain a high-resolution woodland change detection training set and a high-resolution woodland change detection test set;
the method comprises the steps that based on an image alignment module, an existing woodland change detection data set is respectively subjected to image alignment with a high-order woodland change detection training set and a high-order woodland change detection test set, so that an aligned existing woodland change detection data set, an aligned high-order woodland change detection training set and an aligned high-order woodland change detection test set are obtained;
training a deep learning model based on the aligned existing forest land change detection data set, adjusting the deep learning model according to the aligned high-resolution forest land change detection training set, and carrying out change detection on the aligned high-resolution forest land change detection test set according to the adjusted deep learning model to obtain mining area forest land change distribution data;
the objective function of the image alignment module is:
wherein T represents an image converter, D represents an image discriminator, U represents a change detection dataset from a woodland, V represents a change detection training set for a high-resolution woodland or a change detection test set for a high-resolution woodland, D (T (x)) represents a discrimination result of the image discriminator on the image T (x) after the conversion of the image converter, P represents a probability, a first term after the equation represents a probability that the image discriminator discriminates an image converted from an existing dataset into an image of the existing dataset, and a second term represents a probability that the image discriminator discriminates the image converted from the high-resolution dataset into an image of the high-resolution dataset.
2. The method of claim 1, wherein preprocessing the high score satellite image dataset comprises:
carrying out radiometric calibration and geometric correction, single-view image fusion, geometric fine correction, image mosaic, image registration, image clipping, image pair generation, note making and data division on the high-resolution seventh satellite image data.
3. The method of claim 1, wherein the image alignment module comprises an image converter and an image discriminator, the image converter comprising 3 2-dimensional convolution layers, the image discriminator comprising 2-dimensional convolution layers and 2 fully-connected layers.
4. The method for monitoring forest land changes in mining areas based on high-resolution imaging of claim 1, further comprising data normalization of existing forest land change detection data sets, high-resolution forest land change detection training sets, and high-resolution forest land change detection test sets.
5. The method for monitoring the forest land change of the mining area based on the high-resolution image according to claim 1, wherein the deep learning model is a twin U-Net model and comprises an encoder and a decoder, the encoder comprises 5 submodules, each submodule comprises 2 convolution layers with the convolution kernel size of 3×3, and downsampling is carried out on the convolution layers of the first 4 submodules through 1 largest pooling layer;
the decoder comprises 5 sub-modules, each sub-module comprises 2 convolution layers with convolution kernel size of 1×1, and the back 4 sub-modules each perform up-sampling through 1 pooling layer after convolution layers.
6. The method for monitoring forest land changes in mining areas based on high-resolution imaging of claim 5, wherein training the deep learning model based on the aligned existing forest land change detection data set specifically comprises:
carrying out convolution and downsampling operations on the aligned pre-woodland time-phase image and the aligned post-woodland time-phase image for 2 times through a first layer submodule of an encoder respectively, and fusing the pre-time-phase image features and the post-time-phase image features extracted by the first layer;
respectively carrying out convolution and downsampling operations for 2 times on the pre-time-phase image features and the post-time-phase image features extracted by the first layer submodule through the second layer submodule, and fusing the pre-time-phase image features and the post-time-phase image features extracted by the second layer submodule; and so on, until proceeding to the fifth layer sub-module;
inputting the fusion result of the front time phase image feature and the rear time phase image feature output by the fifth layer sub-module of the encoder into the fifth layer sub-module of the decoder, and inputting the fusion result into the fourth layer sub-module of the decoder after convolution operation and up-sampling;
and the fourth layer sub-module of the decoder combines the output result of the fifth layer with the fusion result of the front time phase image feature and the rear time phase image feature extracted by the same layer of the encoder, and inputs the result to the third layer sub-module of the decoder after convolution operation and up-sampling, and the like until the result reaches the first layer sub-module, and outputs a woodland change detection prediction result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310082173.1A CN115984702B (en) | 2023-02-08 | 2023-02-08 | Method for monitoring forest land change of mining area based on high-resolution images |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310082173.1A CN115984702B (en) | 2023-02-08 | 2023-02-08 | Method for monitoring forest land change of mining area based on high-resolution images |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115984702A CN115984702A (en) | 2023-04-18 |
| CN115984702B true CN115984702B (en) | 2023-07-04 |
Family
ID=85972422
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310082173.1A Active CN115984702B (en) | 2023-02-08 | 2023-02-08 | Method for monitoring forest land change of mining area based on high-resolution images |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115984702B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113344130A (en) * | 2021-06-30 | 2021-09-03 | 广州市河涌监测中心 | Method and device for generating differentiated river patrol strategy |
| CN113378727A (en) * | 2021-06-16 | 2021-09-10 | 武汉大学 | Remote sensing image binary change detection method based on characteristic deviation alignment |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| FI20185930A (en) * | 2018-11-02 | 2020-05-03 | Collectivecrunch Oy | Forest stand target attribute prediction |
| CN112183416B (en) * | 2020-09-30 | 2024-02-20 | 北京吉威数源信息技术有限公司 | Automatic extraction method for newly-added construction land based on deep learning method |
| CN114166842A (en) * | 2021-11-19 | 2022-03-11 | 中国自然资源航空物探遥感中心 | Town forest monitoring method based on cooperation of high-resolution remote sensing data and ground survey data |
| CN114782763B (en) * | 2022-05-19 | 2022-11-11 | 哈尔滨工业大学 | Sparse principal component alignment method for multi-view high-resolution remote sensing image |
| CN115527121A (en) * | 2022-10-18 | 2022-12-27 | 国家林业和草原局中南调查规划院 | End-to-end forest land change detection method, system, medium, equipment and terminal |
-
2023
- 2023-02-08 CN CN202310082173.1A patent/CN115984702B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113378727A (en) * | 2021-06-16 | 2021-09-10 | 武汉大学 | Remote sensing image binary change detection method based on characteristic deviation alignment |
| CN113344130A (en) * | 2021-06-30 | 2021-09-03 | 广州市河涌监测中心 | Method and device for generating differentiated river patrol strategy |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115984702A (en) | 2023-04-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111986099B (en) | Tillage monitoring method and system based on convolutional neural network with residual error correction fused | |
| CN103400151B (en) | The optical remote sensing image of integration and GIS autoregistration and Clean water withdraw method | |
| CN111985543A (en) | Construction method, classification method and system of hyperspectral image classification model | |
| CN103914678B (en) | Abandoned land remote sensing recognition method based on texture and vegetation indexes | |
| Xu et al. | Object‐based mapping of karst rocky desertification using a support vector machine | |
| CN113433075B (en) | Simple remote sensing rapid monitoring method for black and odorous water body | |
| CN110334623B (en) | Method for extracting collapsing information based on Sentinel-2A satellite remote sensing image | |
| CN102800074A (en) | Synthetic aperture radar (SAR) image change detection difference chart generation method based on contourlet transform | |
| CN114022783A (en) | Satellite image-based water and soil conservation ecological function remote sensing monitoring method and device | |
| CN116844053B (en) | Wheat planting area identification method, system, electronic equipment and storage medium | |
| CN115861788A (en) | Cultivated land plot extraction method based on high-resolution satellite remote sensing image | |
| CN117522738A (en) | Image cloud removing method and system based on optical remote sensing image and SAR image | |
| CN116740121A (en) | Straw image segmentation method based on special neural network and image preprocessing | |
| CN117422619A (en) | Training method of image reconstruction model, image reconstruction method, device and equipment | |
| CN117075138A (en) | Remote sensing measurement and calculation method, system and medium for canopy height of 30-meter forest in area | |
| CN112949657B (en) | Forest land distribution extraction method and device based on remote sensing image texture features | |
| CN106971402B (en) | SAR image change detection method based on optical assistance | |
| CN117455939A (en) | Cultivated land boundary extraction method of remote sensing satellite image | |
| CN115984702B (en) | Method for monitoring forest land change of mining area based on high-resolution images | |
| CN117557897A (en) | Lodging monitoring method and device for target crops, electronic equipment and storage medium | |
| CN107832805B (en) | Technology for eliminating influence of spatial position error on remote sensing soft classification precision evaluation based on probability position model | |
| CN113657275B (en) | Automatic detection method for forest and grass fire points | |
| CN115100091A (en) | Conversion method and device for converting SAR image into optical image | |
| CN115512159A (en) | Object-oriented high-resolution remote sensing image earth surface coverage classification method and system | |
| CN111768101B (en) | Remote sensing cultivated land change detection method and system taking account of physical characteristics |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |