Land degradation degree classification method based on convolutional neural network and multi-source data

Abstract

The application provides a land degradation degree classification method based on a convolutional neural network and multi-source data, and comprises the following steps: S1, determining a land degradation degree standard; S2, acquiring land degradation degree sample point data and classifying the sample points according to the land degradation degree standard; S3, acquiring environmental variable data and performing preprocessing; S4, uniformly processing the classified land degradation degree sample point data and the preprocessed environmental variable data to obtain sample data; S5, constructing a land degradation degree classification model and training the model; and S6, acquiring a land degradation degree classification result by using the trained land degradation degree classification model. The method can solve the problems in the prior art, such as only focusing on a single environmental index, ignoring other multi-source environmental factors, neglecting complex nonlinear relationships existing in actual processes, resulting in a large deviation of results, and having great subjectivity.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of resource environment and land management, and particularly relates to a land degradation degree classification method based on a convolutional neural network and multi-source data. BACKGROUND

[0002] Land degradation is a global environmental problem that poses a serious threat to ecosystem services, food security and human well-being. The United Nations Sustainable Development Goals (SDGs) have listed "zero growth in land degradation by 2030" as one of the important goals. However, with the increasing population, urbanization, climate warming and irrational use of land resources, the scale and rate of land degradation are increasing, posing serious challenges to agricultural production, ecological security and sustainable economic and social development. Therefore, it is necessary to objectively classify the degree of land degradation and develop accurate land degradation degree classification standards.

[0003] The land degradation degree classification method has undergone a development process from qualitative to quantitative and from single index to multi-index comprehensive. Early land degradation classification mainly relied on expert experience and field investigation. Although this method can obtain accurate point information, it is time-consuming and labor-intensive and is easily affected by human accessibility, terrain, weather and time cost, so it cannot quickly obtain distributed information of land degradation degree in a large area. With the development of earth observation technology, remote sensing technology is used for land degradation monitoring and classification. Based on this, the commonly used land degradation degree classification methods include single index method, empirical model method, comprehensive index method and land use change method. The single index method uses a single environmental index such as normalized difference vegetation index (NDVI), vegetation coverage, land surface temperature, soil organic matter content, etc. to reflect the degree of land degradation. The empirical model method is based on empirical knowledge such as the soil and water loss equation (RUSLE) and the ecological environment quality index (RSEI), and uses soil and water loss and ecological quality to measure the degree of land degradation. The comprehensive index method uses expert scoring method (AHP method, analytic hierarchy process) or objective weighting (entropy weight), and through the weighted comprehensive of multiple environmental indexes (such as vegetation index, soil moisture, rainfall, etc.) to classify the degree of land degradation. The land use change method is based on land use / land cover change (LUCC) to study land degradation, and analyzes the changes of agricultural land, forest land, grassland, wasteland, etc. to judge the degree of land degradation.

[0004] The existing land degradation classification method provides an important theoretical and practical basis for solving the problem of land degradation in early research. Although the above classification method promotes the progress of the land degradation degree classification method, there are still some defects that cannot be ignored: 1) only focus on a single environmental index, ignoring the comprehensive influence of other multi-source environmental factors such as soil properties, topography, climate and other elements; 2) usually assume that the parameters such as rainfall, soil and other parameters affecting land degradation change linearly, ignoring the complex nonlinear relationship in the actual process, resulting in large deviation of the results; 3) the division standard and classification process of land degradation degree still have great subjectivity, and there is no unified, cross-regional comparison of degradation degree classification standard and technical process.

[0005] In order to solve the above problems, it is urgent to develop a new land degradation degree classification method which is objective, universal, multi-source data fusion and nonlinear modeling, in order to cope with the increasingly complex land degradation problem and improve the classification accuracy and efficiency. SUMMARY

[0006] The purpose of the embodiment of the application is to provide a land degradation degree classification method based on convolutional neural network and multi-source data, in order to solve the problems in the prior art that only focus on a single environmental index, ignore other multi-source environmental factors, ignore the complex nonlinear relationship in the actual process, resulting in large deviation of the results, and have great subjectivity.

[0007] To solve the above technical problems, the technical scheme adopted by the application is,

[0008] The land degradation degree classification method based on convolutional neural network and multi-source data comprises the following steps:

[0009] S1: determining the land degradation degree standard;

[0010] S2: obtaining land degradation degree sample data and classifying the sample points according to the land degradation degree standard;

[0011] S3: obtaining environmental variable data and preprocessing;

[0012] S4: uniformly processing the classified land degradation degree sample data and the preprocessed environmental variable data to obtain sample data;

[0013] S5: constructing a land degradation degree classification model and training the model;

[0014] S6: obtaining the land degradation degree classification result using the trained land degradation degree classification model.

[0015] Further, the S1 comprises:

[0016] The land degradation degree standard is divided into:

[0017] No degradation: the soil profile layer structure is complete ABCR layer structure,

[0018] Mild degradation: in the soil profile layer structure, the A layer is completely lost, and the B layer is exposed,

[0019] Moderate degradation: in the soil profile layer structure, A and B layers are completely lost, and C layer is exposed,

[0020] Strong degradation: in the soil profile layer structure, A, B and C layers are lost, and R layer is exposed.

[0021] Further, the S2 comprises:

[0022] S21: Collect soil stripping degree data of n land degradation degree sample points, and record the spatial coordinates of the sample points;

[0023] S22: According to the soil stripping degree data of the sample points, combined with the land degradation degree standard, the land degradation degree of each sample point is classified.

[0024] Further, the S3 comprises:

[0025] S31: Obtain Sentinel-2 data and pre-process:

[0026] Obtain the top-of-atmosphere apparent reflectance TOA data of Sentinel-2; perform median synthesis processing on the Sentinel-2 remote sensing data set; then calculate the spectral index, including: soil-adjusted vegetation index SAVI, normalized difference vegetation index NDVI, enhanced vegetation index EVI, total greenness index SGI, normalized burn ratio NBR, leaf area index LAI, clay mineral index CMR, iron mineral ratio index, iron oxide ratio index, regional burn index;

[0027] S32: Obtain Sentinel-1 data and pre-process:

[0028] Obtain Sentinel-1 ground range detection GRD data, and perform median synthesis; the ground range detection GRD data contains the backscattering coefficient in the vertical emission and vertical reception polarization mode, that is, the VV backscattering coefficient, the backscattering coefficient in the vertical emission and horizontal reception polarization mode, that is, the VH backscattering coefficient; then difference normalization is performed on the VV backscattering coefficient and the VH backscattering coefficient to obtain the radar vegetation index RVI;

[0029] S33: Obtain digital elevation model DEM data and pre-process:

[0030] Obtain DEM data, perform terrain analysis on the DEM data, and obtain parameters such as elevation, slope, slope direction, terrain curvature, terrain slope position index, ground relief degree, runoff dynamic index, ground roughness, terrain humidity index, and slope length.

[0031] Further, the S3 further comprises:

[0032] S34: Obtain soil data, climate data, and land use type data and perform preprocessing:

[0033] Obtain soil data, including soil organic matter content, soil sand content, soil silt content, and soil clay content data; climate data is data of annual average temperature and annual average precipitation during the past a years; obtain land use type data;

[0034] The data obtained in S31-S34 are all environmental variable data, and the forms of the environmental variable data are all grid forms;

[0035] S35: Data gridding;

[0036] Obtain soil parent material type data and geological type data and perform gridding;

[0037] S36: Data standardization;

[0038] Resample the grid data of all environmental variables to the same resolution; and convert the land use type data, the geological type data, and the soil parent material type grid data into numerical forms, and perform normalization processing on the grid data of other environmental variables except the three environmental variables.

[0039] Further, the S4 comprises:

[0040] After unifying the gridded environmental variable data and the sample land degradation degree classification data obtained in S2 to a projection coordinate system, extract all the environmental variable data corresponding to each sample point position, to obtain sample data with a dimension of m*n, where m is the sum of the number of environmental variables and the number of land degradation degree variables, and n is the number of sample points.

[0041] Further, the S5 comprises:

[0042] S51: Encode the land degradation degree variable into a vector form, and preset a model learning rate, a batch, and an iteration number:

[0043] No degradation is [1, 0, 0, 0], slight degradation is [0, 1, 0, 0], moderate degradation is [0, 0, 1, 0], and severe degradation is [0, 0, 0, 1];

[0044] S52: Construct a model and train the model;

[0045] The land degradation degree classification model is composed of an input layer, an output layer, a convolution layer, a pooling layer, a full connection layer, a flattening layer and a Dropout layer.

[0046] In the input layer, the input image size is 64x64, and each pixel point has 32 channels.

[0047] In each convolution layer, the size of the convolution kernel is 3x3, and the number of convolution kernels is 64, 128 and 256.

[0048] In each pooling layer, 2x2 maximum pooling is used.

[0049] In the output layer, a Softmax activation function is used to output the land degradation degree classification result.

[0050] The sample data obtained by S4 is divided into a training set and a validation set, and the training set, a cross-entropy loss function and an Adam optimizer are used to train the model.

[0051] The beneficial effects of the present application are:

[0052] 1. Innovation and objectivity of land degradation degree definition. The present application defines the land degradation degree as the relative stage of land degradation development, and uses the soil profile structure (A layer, B layer, C layer and R layer) exposed after the degradation of the original soil body as the intuitive representation. This method has the following advantages: 1) objectivity: based on observable soil profile structure, reduces subjective judgment, and has clear physical meaning; 2) comparability: provides a unified standard for land degradation degree classification in different regions and different periods; 3) clear physical meaning: intuitively reflects the actual degree of soil degradation, easy to understand and apply; 4) helps to establish a global unified land degradation classification system, improves the comparability and reliability of the classification results.

[0053] 2. Comprehensive and systematic multi-source data fusion. The present application comprehensively utilizes multi-source spectral optical remote sensing, synthetic aperture radar and DEM data obtained from satellite platforms, and climate, soil and geological information obtained from ground monitoring systems as influencing factors of land degradation degree. This multi-source data fusion method has the following advantages: 1) comprehensiveness: covers various factors affecting land degradation. 2) complementarity: different data sources complement each other, providing more comprehensive surface information. 3) multi-scale: combines large-scale observation of satellite remote sensing and fine observation of ground monitoring, thereby improving the accuracy and reliability of classification.

[0054] 3. The innovation and efficiency of the deep learning algorithm. The present application adopts an innovative convolutional neural network architecture, which has the following advantages: 1) automatic feature extraction: deep learning can automatically learn complex feature representation, reducing manual feature engineering. 2) Nonlinear modeling capability: can capture the complex nonlinear relationship in land degradation process. 3) End-to-end learning: from raw data to the final degradation degree, reducing the cumbersome intermediate steps. Significantly improve the accuracy and efficiency of land degradation degree classification, especially in large-scale applications. 4) High precision: deep learning algorithm can effectively capture complex nonlinear relationships and improve classification accuracy.

[0055] 4. The method of the present application designs an automatic classification process, which is suitable for land degradation classification in large-scale areas. This feature has the following advantages: 1) Efficiency: automatic processing of large amounts of data significantly improves classification efficiency. 2) Consistency: uniform classification standards and methods ensure the consistency of the results. 3) Generalizability: the method is universal and can be applied to different regions. It provides a feasible technical solution for regional and even global land degradation monitoring and classification.

[0056] In summary, the method of the present application has innovations and breakthroughs in land degradation degree classification by using multi-source data fusion and deep learning algorithm, significantly improving the scientificity, accuracy and practicability of land degradation classification. These advantages and positive effects make the method have important application value in achieving the goal of zero growth of land degradation and promoting regional sustainable development. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0058] Figure 1 is the process architecture diagram of the present application.

[0059] Figure 2 is the land degradation degree standard diagram.

[0060] Figure 3 is the land degradation degree investigation sample distribution diagram.

[0061] Figure 4 is the land degradation environmental factor importance ranking diagram of the case area of the present application.

[0062] Figure 5 is the land degradation degree classification result distribution diagram of the case area of the present application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0064] The complete steps of the method of the present application are as follows:

[0065] Referring to Figure 1 The steps of the embodiments of the present application are as follows:

[0066] S1: Determine the land degradation degree standard.

[0067] Referring to Figure 2 The present application firstly defines the land degradation degree as the relative stage of land degradation development, and takes the current soil profile layer structure (A layer, B layer, C layer and R layer) exposed after the original soil body is degraded as the intuitive representation. Specifically, it is divided into four degradation degrees,

[0068] No degradation: complete ABCR layer structure,

[0069] Mild degradation: A layer is completely lost, and B layer is exposed,

[0070] Moderate degradation: B layer is completely lost, and C layer is exposed in large area,

[0071] Severe degradation: most of C layer is lost, and R layer is exposed.

[0072] The specific description of each degradation degree is shown in Table 1,

[0073] Table 1: Land degradation degree and characteristic description

[0074]

[0075] This method can be used to obtain the land degradation degree through field investigation. This definition method is based on the observable soil profile structure, reduces subjective judgment, provides a unified standard for land degradation degree classification in different regions and different periods, intuitively reflects the actual degree of soil degradation, is convenient for understanding and application, and is helpful for establishing a global unified land degradation classification system and improving the comparability and reliability of the classification results.

[0076] S2: Land degradation degree sample data acquisition and sample classification. Through field investigation, the sample data of land degradation degree in the case area were obtained, and the land degradation degree of each sample point was classified by observing the soil stripping degree combined with the classification standard of land degradation degree proposed in S1. The determined land degradation degree was used as the dependent variable in the land degradation degree classification model. The specific steps are as follows:

[0077] S21: Refer to Figure 3 Through field investigation, 297 sample data of land degradation degree in the case area in 2020 were collected, i.e. the soil stripping degree information of sample points, and the spatial coordinate information of sample points was recorded by using handheld GPS instrument.

[0078] S22: According to the soil stripping degree information of sample points in field investigation, combined with the proposed classification standard of land degradation degree, the land degradation degree of each sample point was classified, and the land degradation degree of sample was used as the dependent variable of the model.

[0079] S3: Acquisition and processing of land degradation degree influencing factors, i.e. environmental data.

[0080] Since land degradation is caused by the combined action of natural and human activities, the multispectral remote sensing data (Sentinel-2 satellite), synthetic aperture radar (Sentinel-1 satellite), DEM, soil properties, climate, land use, and geological type data of the case area at the same period were obtained as the independent variables of the land degradation degree classification model. The specific steps are as follows:

[0081] After preprocessing these environmental data, the independent variables of the model were obtained as shown in Table 2.

[0082] Table 2 Independent variables used in land degradation degree classification model

[0083]

[0084]

[0085] S31: Acquisition and processing of Sentinel-2 data.

[0086] Get all Sentinel-2 top-of-atmosphere apparent reflectance TOA product data with spatial resolution of 10, 20, 60 m in the case area in 2020 on the Google earth engine remote sensing cloud platform; the data has been processed by radiation calibration, atmospheric correction and geometric correction; the cloud cover of the image is counted, and only the data with cloud cover less than 5% is retained; the median synthesis processing is performed on all Sentinel-2 remote sensing data sets that meet the conditions to reduce cloud cover and data noise; then 10 spectral indices are extracted by band calculation, including soil adjusted vegetation index (SAVI), normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), sum green index (SGI), normalized burn ratio (NBR), leaf area index (LAI), clay minerals ratio (CMR), ferrous minerals ratio (FMR), iron oxide ratio (IOR), and burn area index (BAI). These spectral indices can retrieve land cover, vegetation vigor, soil characteristics, and roughness.

[0087] S32: Acquisition and processing of Sentinel-1 data.

[0088] Get all Sentinel-1 GRD product data in the case area in 2020 on the Google earth engine remote sensing cloud platform after spot filtering and terrain correction processing; it contains backscattering coefficients in VV and VH polarization modes; all Sentinel-1 synthetic aperture radar data are median synthesized to reduce noise and obtain clearer images, and then the backscattering coefficients in VV and VH polarization modes are normalized and calculated to obtain radar vegetation index (RVI).

[0089] S33: Processing of DEM data.

[0090] The STRM DEM v3.0 product DEM data with a spatial resolution of 30 m in the case area is downloaded through the Google earth engine remote sensing cloud platform, and then the regional DEM data is subjected to terrain analysis by using SAGA GIS software, so as to extract the terrain parameters such as elevation, slope, aspect, terrain curvature, terrain position index, surface relief, runoff power index, surface roughness, terrain humidity index and slope length.

[0091] S34: Acquisition and processing of soil, climate data, and land use type.

[0092] The 90m spatial resolution soil property raster data of the place is obtained from the Resource and Environment Science and Data Center of the Chinese Academy of Sciences, including soil organic matter, sand, silt and clay content data; the meteorological data is the 1km×1km raster data of annual average temperature and annual average precipitation for many years, which needs to reflect the characteristics of the climate background, and the longer the time series is, the better, in this embodiment, the 1km×1km raster data of annual average temperature and annual average precipitation for 43 years is selected;

[0093] Land use type raster data is obtained under 30m spatial resolution;

[0094] The above environmental variable data are in the form of raster data,

[0095] S35: Data rasterization.

[0096] Obtain soil parent material type data; obtain the geological type data of the region from the China Geological Survey Center, with a scale of 1:200,000.

[0097] Through the ArcGIS10.8 platform, the soil parent material type and geological type vector geological type data obtained are rasterized.

[0098] S36: Data standardization.

[0099] In the ArcGIS10.2 software, the raster data of all 32 environmental variables is resampled to 30m spatial resolution by using the cubic method; in addition, except for the land use, geological type and soil parent material type raster data which are subjected to dummy variable processing and converted into numerical form, the other environmental variables are subjected to normalization processing.

[0100] S4: Construction of sample data.

[0101] The 32 environmental variable data (raster spatial data) and the classified land degradation degree sample data (spatial point data) are unified to the UTM-WGS-1984 projection coordinate system by using ArcGIS 10.8 software, and the corresponding environmental variable information of each sample point position is extracted, thereby obtaining sample data with a dimension of 33*297, wherein 33 is the number of variables, and 297 is the number of samples.

[0102] S5: Model construction and precision classification.

[0103] The 297 land degradation degree samples in the case area are used as the dependent variable, and the spectral remote sensing data (Sentinel-2 satellite), synthetic aperture radar (Sentinel-1 satellite), DEM, soil properties, climate, land use, and geological type data in S36 after standardization are used as the independent variable. A land degradation degree classification model based on convolutional neural network is first constructed on the Python platform, and the accuracy of the classification model and the classification results is quantitatively evaluated. The specific steps include the following steps:

[0104] S51: Land degradation degree digital expression and presetting model learning rate, batch, and iteration number.

[0105] The land use degradation degree is encoded as a one-hot vector, wherein no degradation is [1, 0, 0, 0], mild degradation is [0, 1, 0, 0], moderate degradation is [0, 0, 1, 0], and severe degradation is [0, 0, 0, 1]. This facilitates the operation of the convolutional neural network algorithm.

[0106] In the model training process, the cross-entropy loss function and the Adam optimizer are used, the initial learning rate is 0.001, the batch size is 32, the iteration number is 200, the learning rate scheduling strategy is to reduce the learning rate by half every 50 epochs, and the early stopping strategy is to stop training if the loss does not improve within 20 epochs.

[0107] S52: Constructing and training the model.

[0108] The obtained sample data with a dimension of 33*297 is randomly divided into a training set (209) and a validation set (88) in a ratio of 7:3, which are used for model training and model precision verification, respectively.

[0109] Based on the data structure characteristics of the present embodiment, a convolutional neural network classification model is constructed, and the structure is as shown in Table 3.

[0110] Table 3: Convolutional neural network classification model parameters of the present embodiment

[0111]

[0112]

[0113] The land degradation degree classification model based on the convolutional neural network is trained using the training set, fully utilizes the advantages of deep learning, can automatically extract features from multi-source data, establishes complex nonlinear relationships, and realizes accurate and objective land degradation degree classification.

[0114] S6: Obtain the predictive classification result of the land degradation degree.

[0115] The land degradation degree classification model based on the convolutional neural network that has been trained in S5 is applied to the 32 environmental variable raster data sets of the case area with a spatial resolution of 30m and classification is performed, so as to obtain the predictive classification result of the land degradation degree of the case area in 2020, as shown in FIG. 6. Figure 5 .

[0116] Model accuracy:

[0117] The model trained in S52 is verified using the verification set. The overall accuracy (OA), kappa coefficient (k), producer accuracy (PA), and user accuracy (UA) are used to quantify and evaluate the reliability of the model. The OA is a relatively comprehensive index, which is represented as the total number of correctly classified samples divided by the total number of verification samples:

[0118]

[0119] where N is the total number of verification samples, i refers to each class in the confusion matrix, in the confusion matrix, each class i corresponds to a row and a column of the matrix, r is the total number of classes, x ii is the number of true positives of each class, i.e. the number of samples with the same actual and classified categories, x +i represents the total number of samples classified as this class, i.e. the sum of the ith row; and x i+ represents the total number of samples actually belonging to this class, i.e. the sum of the ith column.

[0120] Referring to Table 4, the accuracy OA of the land degradation degree classification model integrating multi-source data and convolutional neural network is 82.02%, and the Kappa is 73.30%.

[0121] Table 4: Land degradation degree confusion matrix (N = 89)

[0122]

[0123]

[0124] Figure 5In the case, the classification results provide a comprehensive analysis of the land degradation status of the case area. The distribution of land degradation degree: 45.03% of the land is slightly degraded, 36.31% is moderately degraded, 9.54% is severely degraded, and 9.12% is extremely severely degraded. Spatial distribution characteristics: the central mountainous area, i.e. the valley area, is mainly non-degraded and slightly degraded; moderate and severe degradation occurs on gentle slopes and flat hills, and extremely severe degradation is mainly distributed on mountain ridges and slope tops with poor vegetation growth and gentle surface.

[0125] Based on the trained convolutional neural network-based land degradation degree classification model in S5, the importance of a single environmental variable to the model is calculated using the Permutation Importance method, which can be represented as the importance ranking of the land degradation impact factor in the case area.

[0126] The importance scores of all environmental variables are normalized to the range of 0-1 for comparison, and the results are as follows Figure 4 .

[0127] Figure 4 The normalized difference vegetation index, soil organic matter content, multi-year average precipitation, leaf area index, and radar vegetation index show high importance, with importance of 15.06%, 13.90%, 12.10%, 11.88%, and 10.59%, respectively; reflecting the important indicative role of vegetation cover and soil quality on land degradation status. This analysis provides a comprehensive perspective to understand the relative importance of different environmental factors in the process of land degradation. Not only can it guide us to use these variables more targeted in the model, but also provide scientific basis for land management and ecological restoration.

[0128] Each embodiment in the specification is described in a related manner, and the same or similar parts between each embodiment can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

[0129] The above only describes the preferred embodiments of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application is included in the protection scope of the present application.

Claims

1. A land degradation classification method based on convolutional neural networks and multi-source data, characterized in that, Includes the following steps: S1: Determine the standards for the degree of land degradation; S2: Obtain sample data on the degree of land degradation and classify the samples according to the standards for the degree of land degradation; S3: Obtain environment variable data and preprocess it; S4: The classified land degradation sample data and the preprocessed environmental variable data are processed in a unified manner to obtain sample data; S5: Construct a land degradation classification model and train the model; S6: Use the trained land degradation classification model to obtain the land degradation classification results; S3 includes: S31: Obtain Sentinel-2 data and perform preprocessing: Get Apparent reflectance at the top of the atmosphere Data; The remote sensing dataset undergoes median synthesis processing; then spectral indices are calculated, including the soil-modified vegetation index. Normalized Difference Vegetation Index Enhanced vegetation index Total Greenness Index Normalized combustion ratio Leaf area index Clay mineral index Iron ore ratio index, iron oxide ratio index, regional combustion index; S32: Acquire Data and preprocessing: Get Ground-range detection GRD data is obtained and median synthesis is performed. The ground-range detection GRD data includes the backscattering coefficients (VV backscattering coefficient) under vertical transmission and vertical reception polarization modes, and the backscattering coefficients (VH backscattering coefficient) under vertical transmission and horizontal reception polarization modes. The radar vegetation index (RVI) is obtained by differential normalization of the VV and VH backscattering coefficients. S33: Acquire and preprocess Digital Elevation Model (DEM) data: Obtain DEM data, perform terrain analysis on the DEM data, and obtain parameters such as elevation, slope, aspect, terrain curvature, terrain slope index, surface relief, runoff dynamic index, surface roughness, terrain humidity index, and slope length. S34: Acquire soil data, climate data, and land use type data, and perform preprocessing: Obtain soil data, including soil organic matter content, soil sand content, soil silt content, and soil clay content; climate data, including the annual average temperature and annual average precipitation from the current year to a years ago; and obtain land use type data. The data obtained in S31~S34 are all environmental variable data, and these environmental variable data are all in raster format. S35: Data rasterization; The acquired soil parent material type data and geological type data are rasterized; S36: Data standardization; All environmental variable raster data were resampled to the same resolution; land use type data, geological type data, and soil parent material type raster data were converted into numerical form, and the raster data of other environmental variables were normalized.

2. The land degradation classification method based on convolutional neural networks and multi-source data according to claim 1, characterized in that, S1 includes: The degree of land degradation is classified according to the hierarchical structure of soil profiles: No degradation: The soil profile has a complete ABCR layer structure. Mild degradation: In the soil profile, layer A is completely lost, and layer B is exposed. Moderate degradation: In the soil profile, layers A and B are completely lost, and layer C is exposed. Strength degradation: In the soil profile layer structure, layers A, B, and C are lost, and layer R is exposed.

3. The land degradation classification method based on convolutional neural networks and multi-source data according to claim 2, characterized in that, S2 includes: S21: Collect soil stripping data from n land degradation sampling points and record the spatial coordinates of the sampling points; S22: Based on the soil stripping degree data of the sample points and combined with the land degradation degree standards, classify the land degradation degree of each sample point.

4. The land degradation classification method based on convolutional neural networks and multi-source data according to claim 1 or 3, characterized in that, S4 includes: After unifying the rasterized environmental variable data and the land degradation classification data of the sample points obtained in S2 into the projected coordinate system, all environmental variable data corresponding to each sample point location are extracted to obtain the dimension of The sample data, of which The number of variables is the sum of the number of environmental variables and the number of variables related to the degree of land degradation. The number of sample points.

5. The land degradation classification method based on convolutional neural networks and multi-source data according to claim 4, characterized in that, S5 includes: S51: Encode the land degradation degree variable into vector form, and preset the model learning rate, batch size, and number of iterations: No degradation is represented by [1, 0, 0, 0], slight degradation by [0, 1, 0, 0], moderate degradation by [0, 0, 1, 0], and severe degradation by [0, 0, 0, 1]; S52: Build and train the model; The land degradation classification model consists of an input layer, an output layer, a convolutional layer, a pooling layer, a fully connected layer, a flattening layer, and a dropout layer. In the input layer, the input image size is 64 x 64, and each pixel has 32 channels; In each convolutional layer, the kernel size is 3x3, and the number of kernels is 64, 128, or 256. In each pooling layer, 2x2 max pooling is used; In the output layer, the Softmax activation function is used to output the classification results of land degradation degree; The sample data obtained from S4 is divided into a training set and a validation set. The model is trained using the training set, the cross-entropy loss function, and the Adam optimizer.

Images