CN113780105B

CN113780105B - Space-time mixed pixel decomposition method for MODIS real-time remote sensing image data

Info

Publication number: CN113780105B
Application number: CN202110972899.3A
Authority: CN
Inventors: 王群明; 丁欣宇
Original assignee: Tongji University
Current assignee: Tongji University
Priority date: 2021-08-24
Filing date: 2021-08-24
Publication date: 2023-07-07
Anticipated expiration: 2041-08-24
Also published as: CN113780105A

Abstract

The invention relates to a space-time mixed pixel decomposition method for MODIS real-time remote sensing image data, which comprises the steps of 1) obtaining a scene MODIS-Landsat image pair which covers the same area with an MODIS image to be mixed pixel decomposed and is closest in time as auxiliary data; 2) Judging whether the image pair has cloud pollution or not, if not, directly calculating the spectrum difference value between the MODIS data at two moments; if yes, only calculating the MODIS spectrum difference value of the completely cloud interference free region in the two moments; extracting MODIS pixels which are not subjected to ground object coverage change in two times; 3) Directly combining spectrum of the MODIS pixels extracted without cloud pollution with corresponding ground object mixing proportion to construct a training sample; screening and amplifying the spectrum of the MODIS pixels extracted by cloud pollution and the mixing proportion of the corresponding ground objects to form a new training sample; 4) Training a learning model corresponding to the MODIS moment, and estimating the mixing proportion of all the residual variation MODIS pixels. Compared with the prior art, the invention has the advantages of strong practicability, high precision and the like.

Description

Space-time mixed pixel decomposition method for MODIS real-time remote sensing image data

Technical Field

The invention relates to the technical field of remote sensing image processing, in particular to a space-time mixed pixel decomposition method for MODIS real-time remote sensing image data.

Background

The mixed pixel is an inherent problem of the remote sensing image, and the corresponding spectrum is regarded as the mixture of the spectra of the multiple types of ground objects. The mixed pixel decomposition technology is proposed for the mixed pixel problem, is also called soft classification, and is a technology for estimating the proportion of various ground objects in the pixel. Compared with the traditional hard classification technology, the mixed pixel decomposition can break through the limitation of the pixel self dimension, realize sub-pixel level information extraction and provide more abundant ground object information. Therefore, the theory and the method for developing the corresponding mixed pixel decomposition have great significance.

Over the past decade, researchers have successively proposed a series of spectral mixing models to address the problems with mixing pixels, with the more common applications including linear mixing and nonlinear mixing models. The linear mixed model has been widely paid attention to since the mixed pixel decomposition technology is proposed because of its definite physical meaning, simplicity and high efficiency, and the solution process of the mixed proportion is essentially the solution process of the linear coefficient by regarding each mixed pixel as the linear combination of various end members. Most linear hybrid models assume that the spectrum of each type of feature can be represented by a single fixed end member. However, the phenomenon of alien spectrum is commonly existed in remote sensing images, and the spectrum difference in the class is usually large, i.e. the spectrum presented by each pixel will also have a large difference in the same class of ground objects. Therefore, the spectrum of a type of ground object represented by a single element is not accurate. The spectrum difference in the end member is an important obstacle for improving the decomposition precision of the mixed pixels. To overcome the problem of spectrum difference in classes, researchers sequentially put forward a series of multi-end member mixed decomposition methods, and the core idea is to assume that the spectrum of each class of ground object needs to be described by a plurality of end members, including a multi-end member spectrum mixed analysis method, an end member beam method, a Monte Carlo method and the like. While the multi-terminal approach can overcome the problem of intra-class spectral differences to some extent, a large number of terminal members are often required for description. Especially for regions of high heterogeneity, the presence of mixed pixels in large numbers and the extraction of a large number of pure end members often presents a great challenge. In recent years, with the continuous development and intensive research of machine learning and deep learning, methods such as support vector machines, neural networks and the like are widely applied to mixed pixel decomposition. While these methods may not require a clean end member, it is also necessary to obtain a large amount of mixing ratio supervision information to construct training samples.

The existing mixed decomposition model is usually carried out based on single-time data, so that the mixed decomposition problem is large in uncertainty and more in error source. And less research is done on the decomposition of the mixed pixels of the time series data, and the relation of the data in the time domain is ignored. Meanwhile, in practical application, the acquired real-time remote sensing data is very necessary to be analyzed and interpreted in time, so that the change of the ground object can be rapidly judged, and the rapid dynamic change of the ground object can be monitored. Therefore, the space-time mixing decomposition method which can process real-time data and monitor the coverage change of the ground features has very important theoretical significance and application value.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide a space-time mixed pixel decomposition method for MODIS real-time remote sensing image data.

The aim of the invention can be achieved by the following technical scheme:

a method for decomposing spatio-temporal hybrid pixels of MODIS real-time remote sensing image data, the method comprising the steps of:

s1: acquiring a scene MODIS-Landsat image pair which covers the same area with the MODIS image decomposed by the pixels to be mixed and is closest in time as auxiliary data;

s2: judging whether the MODIS-Landsat image pair has cloud pollution or not, if so, directly calculating the spectrum difference value between the MODIS data at two moments; if cloud pollution exists, only calculating the MODIS spectrum difference value of the completely cloud interference free region in two moments; after calculating the spectrum difference value under the two conditions, extracting MODIS pixels which have no ground object coverage change between the two moments through change detection;

further, in S2, a mode of directly subtracting two MODIS remote sensing images is adopted to calculate a mode value, and a spectrum difference value is obtained through the mode value.

Further, methods of extracting the MODIS pixels where the change in the ground cover does not occur between the two moments include, but are not limited to, the OTSU method.

S3: the training sample is constructed by directly combining the spectrum of the MODIS pixels extracted under the condition of S2 without cloud pollution and the corresponding ground object mixing proportion; the spectrum of the MODIS pixel extracted under the cloud pollution condition and the corresponding ground object mixing proportion are screened and amplified to form a new training sample;

further, the ground object mixing proportion corresponding to the extracted MODIS pixels is obtained through degradation of the classification drawing result of Landsat data at the corresponding moment.

Further, methods for training sample augmentation, including but not limited to linear hybrid reconstruction.

The specific contents for screening and amplifying the spectrum of the MODIS pixel extracted under the condition of cloud pollution and the corresponding ground object mixing proportion are as follows:

firstly, screening reliable training samples according to corresponding proportion, and then carrying out average grouping on the screened training samples; then selecting the combination of two pairs and simultaneously carrying out linear mixed reconstruction to be used as an amplified sample; and taking the screened training sample and the amplified sample as new training samples.

The specific content of the reliable training samples screened according to the corresponding proportion is as follows:

and (3) selecting a pixel with a certain proportion of the minimum modulus as a reliable sample according to the modulus obtained by calculation in the step (S2). Further, a certain proportion is preferably 20% -100%.

S4: training a learning model corresponding to the MODIS moment by using the training sample obtained in the step S3 and the new training sample, and estimating the mixing proportion of all the residual varied MODIS pixels by using the trained learning model.

The learning model includes, but is not limited to, a least squares support vector machine model. And taking the MODIS spectrum of the residual pixels at the known time to be decomposed as the input of a prediction sample, inputting a trained learning model, and predicting the mixing proportion of the MODIS pixels of all the residual changes at the time to be decomposed.

Compared with the prior art, the space-time mixed pixel decomposition method for MODIS real-time remote sensing image data provided by the invention at least has the following beneficial effects:

1. the mixed pixel decomposition method provided by the invention can directly extract the supervision information of unchanged pixels, and compared with the traditional mixed decomposition technology, the mixed pixel decomposition method avoids end member extraction and can take into account intra-class spectrum differences.

2. Compared with the prior mixed pixel decomposition technology aiming at a single time point, the method has the advantage of larger error. The invention can extend from the traditional space domain to the time-space domain, and utilizes the MODIS-Landsat image pair as auxiliary data to carry out mixed pixel decomposition so as to improve the mixed decomposition precision.

3. According to the invention, the mixed pixel decomposition technology is expanded from historical data to real-time data processing, and the MODIS real-time remote sensing image data can be used for monitoring real-time surface coverage change and rapidly judging the change of the ground object.

Drawings

FIG. 1 is a flow chart of a method for decomposing spatio-temporal hybrid pixels for MODIS real-time remote sensing image data in an embodiment;

FIG. 2 is a diagram showing the decomposition result of region one in the embodiment;

fig. 3 is a diagram showing the decomposition result of the second region in the embodiment.

Detailed Description

The invention will now be described in detail with reference to the drawings and specific examples. It will be apparent that the described embodiments are some, but not all, embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present invention without making any inventive effort, shall fall within the scope of the present invention.

Examples

The MODIS series data become one of the most commonly used data for global surface monitoring due to the relatively high time resolution, and the Landsat series data have relatively high spatial resolution and are also widely applied to the production of surface coverage products. Based on the requirement of rapid dynamic change monitoring, a method for decomposing mixed pixels of MODIS real-time remote sensing data needs to be researched. Acquisition of pure end members and mixed ratio supervision information often presents a significant challenge due to the low spatial resolution of the MODIS data. However, the time resolution of the MODIS data is high, and the spatial resolution of the corresponding Landsat data is high, so that the complementary advantage can be fully utilized, and the mixed pixel decomposition is expanded from the traditional time domain to the time-space domain. Specifically, the time domain adjacent data MODIS-Landsat image pair of MODIS real-time data to be mixed and decomposed can be used as auxiliary data to extract the mixed proportion supervision information, so that the complex operation of end member extraction is avoided, and the MODIS data mixed proportion with higher precision is obtained.

Based on the thought, the invention relates to a space-time mixed pixel decomposition method for MODIS real-time remote sensing image data, which comprises the following specific steps:

step one, obtaining a scene MODIS data and Landsat data which are closest in time (namely shortest in time interval with the decomposing data of the pixels to be mixed) and cover the same area with the MODIS image decomposed by the pixels to be mixed, and forming an MODIS-Landsat image pair as auxiliary data.

In the remote sensing image, a scene represents a picture taken by a satellite once.

And step two, judging whether the MODIS-Landsat image pair has cloud pollution or not. If no cloud pollution exists, directly calculating the spectrum difference value between MODIS data at two moments (namely the moment of the MODIS data to be decomposed and the moment of the acquired auxiliary data); otherwise, only the MODIS spectrum difference of the completely cloud interference free region in two moments is calculated. The larger the spectral difference, the greater the likelihood of a change in the surface features. And (3) calculating a spectrum difference value under the two conditions, and then carrying out change detection on a spectrum difference value result, so as to extract an MODIS pixel without ground object coverage change between the two moments. Specifically, a mode of directly subtracting two MODIS remote sensing images can be used, and then a mode value of the MODIS remote sensing images is calculated to obtain a spectrum difference value. And further carrying out change detection by using an OTSU algorithm to extract unchanged pixels.

And thirdly, directly combining the spectrum of the MODIS pixels extracted under the cloud pollution-free condition in the second step and the corresponding ground feature mixing proportion (obtained by degradation of the classification drawing result of Landsat data at the corresponding moment) to serve as training samples. On the contrary, the spectrum of the MODIS pixel extracted under the condition of cloud pollution and the corresponding ground object mixing proportion are screened and amplified; as a preferred scheme, more reliable samples are screened according to the corresponding proportion, specifically, according to the modulus value calculated in the second step, a pixel with a certain proportion having the minimum modulus value is selected as the reliable sample, and a certain proportion can be preferentially selected to be 20% -100%, namely 20% -100% of the samples are extracted in the second step. And then carrying out average grouping on the training samples after screening, then selecting two-by-two combinations and simultaneously carrying out linear mixed reconstruction to obtain the amplified samples. The screened sample and the amplified sample are simultaneously used as new training samples.

And step four, training samples formed in the step three under two conditions are used for training a least square support vector machine learning model corresponding to the MODIS moment. And taking the MODIS spectrum of the residual pixels at the known time to be decomposed as the input of a prediction sample, and predicting the mixing proportion of all the residual variation MODIS pixels at the time to be decomposed by using a trained learning model. And finally obtaining the mixing proportion value of the MODIS image decomposed by the pixels to be mixed.

In order to verify the effectiveness of the method, the method is adopted to decompose MODIS remote sensing image data, simulate MODIS-Landsat image data under the condition of cloud pollution, and compare the mixed pixel decomposition result with the mixed decomposition result of the conventional common method, namely a linear mixed decomposition model (marked as LSMM) and a machine learning result (marked as LSSVM) by taking auxiliary MODIS data as training samples. For the case of no cloud pollution of auxiliary data, the selected test area is located in Beijing Daxing area (area one) of China, and the data acquisition time is 2014, 9, 4 and 10, 6. And under the condition that the auxiliary data has cloud pollution, selecting a test area to be positioned in the Wuhan region (area two) in Hubei province of China. The data acquisition time is 5 months, 3 days and 7 months, 22 days in 2001. For both the above-mentioned areas, the present embodiment performs hybrid decomposition on the MODIS data at the later time by using the MODIS and Landsat data at the former time and the MODIS data at the later time. Meanwhile, landsat data (i.e., the MODIS mixing ratio after degradation thereof) at the later moment is used as real reference data for precision evaluation.

As can be seen from fig. 2 to 3 (white portions in the auxiliary data represent cloud pollution), the mixed decomposition result using LSMM shows aggregation of large-scale colors, and accurate prediction of regions having strong spatial heterogeneity is not possible, and the result is greatly different from the reference map. Meanwhile, by using the LSSVM method, all pixels at the previous time point are extracted for training, but due to the ground feature change and the difference between spectrums between the two time points, a training sample extracted based on the data at the previous time point cannot accurately describe the spectrum and mixing proportion relation at the predicted time point, namely the training sample has larger uncertainty and limited machine learning effect. Therefore, the results are also greatly different from the reference figures. Because the strategy of the invention is to extract unchanged pixels according to the MODIS spectrum difference between two time points, namely, to extract a training sample for training based on the prediction moment data, wherein the spectrum of the training sample is closer to that of a test sample. Further, in the case of cloud pollution, the strategy of the invention screens and amplifies samples on the basis of detecting the extracted samples through the change of the cloud-free area, and provides a sufficient number of more reliable training samples. Therefore, the invention can obtain a higher-precision mixed decomposition result, which is closer to a reference result, no matter whether cloud pollution exists or not.

The mixed pixel decomposition results of each method were evaluated for accuracy using correlation coefficients (Correlation Coefficient, CC), root mean square error (Root Mean Square Error, RMSE) and mean absolute error (Mean Absolute Error, MAE) evaluation indices, as shown in table 1. Wherein the CC reflects the correlation degree between the predicted result and the real result, and the larger the value of the CC is, the closer the predicted result and the real result are; the difference between the predicted result and the real result is measured by the RMSE, and the larger the value of the difference is, the more the predicted result deviates from the real result; similarly, the MAE reflects the difference of the predicted outcome from the true outcome, with a larger value indicating that the predicted outcome deviates from the true outcome.

Table 1 evaluation of accuracy of the results of the mixing ratio

As can be seen from the objective evaluation results in Table 1, the method predicts the mixed pixel proportion result of the MODIS real-time remote sensing data, and the mixed pixel proportion result is superior to the mixed pixel proportion result of other two common methods no matter whether cloud pollution exists or not, and all indexes indicate that the strategy can obtain the mixed proportion which is closer to the real situation.

In conclusion, the mixed pixel decomposition strategy has obvious advantages in vision and precision evaluation, the obtained mixed decomposition precision is highest, the mixed decomposition effect is optimal, and the mixed pixel decomposition strategy is a feasible mixed pixel decomposition strategy.

While the invention has been described with reference to certain preferred embodiments, it will be understood by those skilled in the art that various changes and substitutions may be made without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the protection scope of the invention is subject to the protection scope of the claims.

Claims

1. The space-time mixed pixel decomposition method for the MODIS real-time remote sensing image data is characterized by comprising the following steps of:

1) Acquiring a scene MODIS-Landsat image pair which covers the same area with the MODIS image decomposed by the pixels to be mixed and is closest in time as auxiliary data;

2) Judging whether the MODIS-Landsat image pair has cloud pollution or not, if so, directly calculating the spectrum difference value between the MODIS data at two moments; if cloud pollution exists, only calculating the MODIS spectrum difference value of the completely cloud interference free region in two moments; after calculating the spectrum difference value under the two conditions, extracting MODIS pixels which have no ground object coverage change between the two moments through change detection;

3) Directly combining the spectrum of the MODIS pixels extracted under the condition of no cloud pollution in the step 2) with the corresponding ground object mixing proportion to construct a training sample; the spectrum of the MODIS pixel extracted under the cloud pollution condition and the corresponding ground object mixing proportion are screened and amplified to form a new training sample;

4) Training a learning model corresponding to the MODIS moment by using the training sample and the new training sample obtained in the step 3), and estimating the mixing proportion of all the residual variation MODIS pixels by using the trained learning model.

2. The method for decomposing the spatio-temporal hybrid pixels of the MODIS real-time remote sensing image data according to claim 1, wherein a mode of directly subtracting two MODIS remote sensing images is adopted to calculate a mode value, and a spectrum difference value is obtained through the mode value.

3. The method for decomposing the spatio-temporal hybrid pel of the MODIS real-time remote sensing image data according to claim 2, wherein in the step 2), the method for extracting the MODIS pel without the change of the ground object coverage between the two moments includes, but is not limited to, the OTSU method.

4. The method of claim 2, wherein in step 3), the method for training sample augmentation includes, but is not limited to, linear hybrid reconstruction.

5. The method of claim 1, wherein the learning model includes, but is not limited to, a least squares support vector machine model.

6. The space-time mixed pixel decomposition method for MODIS real-time remote sensing image data according to claim 1, wherein the ground object mixing proportion corresponding to the extracted MODIS pixel is obtained by degradation of the classification drawing result of the corresponding moment Landsat data.

7. The method for spatio-temporal mixed pel decomposition of MODIS real-time remote sensing image data according to claim 1, wherein in step 4), the residual pel MODIS spectrum at the known time to be decomposed is used as the input of a prediction sample, and the trained learning model is input to predict the mixing proportion of all the residual varied MODIS pels at the time to be decomposed.

8. The method for decomposing the space-time mixed pixels of the MODIS real-time remote sensing image data according to claim 4, wherein the specific contents for screening and amplifying the spectrum of the MODIS pixels extracted under the condition of cloud pollution and the corresponding ground object mixing proportion are as follows:

9. The method for decomposing the spatio-temporal hybrid pel of the MODIS real-time remote sensing image data according to claim 8, wherein the specific content of the reliable training samples screened out according to the corresponding proportion is:

and D, selecting a pixel with a certain proportion of the minimum modulus as a reliable sample according to the modulus calculated in the step two.

10. The method for decomposing the spatio-temporal hybrid pixel of the MODIS real-time remote sensing image data according to claim 9, wherein a certain proportion is 20% -100%.