A land degradation status assessment method and system

Abstract

The present invention discloses a method and system for assessing land degradation, which relates to the field of image processing technology. The method involves acquiring land images using multi-source remote sensing, classifying land organisms through multi-level extraction of remote sensing spatial features, constructing a spatiotemporal coupled dynamic particle evolution model based on land organism distribution maps, and predicting the dynamic particle evolution of land organisms. A model for quantifying abiotic value is established, which dynamically calculates a multidimensional value indicator system based on weights assigned according to historical data and machine learning algorithms. Land degradation is assessed based on the historical values of abiotic land values and the dynamic particle evolution predictions of land organisms. The method combines deformable convolution with transfer learning for adaptive feature extraction, improving classification accuracy in complex terrain. Land degradation is assessed by constructing a particle evolution model to quantify the spatiotemporal response of biomes and simultaneously constructing a multidimensional abiotic value indicator system, thereby enhancing the credibility of blockchain technology.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method and system for assessing land degradation status. Background Art

[0002] Land degradation encompasses not only soil erosion, salinization, and desertification, but also encompasses multiple aspects, including desertification, biodiversity loss, and soil structural degradation. To effectively assess land degradation and implement appropriate protective measures, traditional assessment methods rely on ground surveys and field sampling. These methods are often costly, time-consuming, and unable to cover large areas. With advances in remote sensing technology, multi-source remote sensing data (such as satellite and drone remote sensing) has become a crucial tool for assessing land degradation. Remote sensing technology can rapidly and accurately obtain large-scale land degradation information by analyzing surface reflection and scattering information in different wavelength bands, supporting the rational management and protection of land resources.

[0003] Currently, a Chinese invention patent application with application number CN2019103876513 discloses a method and system for assessing land degradation by integrating multiple remote sensing indicators. This method estimates the degree of land degradation based on remote sensing, integrates multiple indicators to derive a comprehensive land degradation index, and enables rapid identification of land degradation from remote sensing imagery. However, existing technologies cannot achieve adaptive feature extraction through the combination of deformable convolution and transfer learning, cannot improve classification accuracy in complex terrain, cannot quantify the spatiotemporal response of biological communities by constructing particle evolution models, and cannot simultaneously construct a multidimensional abiotic value indicator system to assess land degradation. Furthermore, these technologies cannot achieve data transparency and support cross-institutional collaborative governance. Summary of the Invention

[0004] The technical problems solved by the present invention are: the existing technology cannot perform adaptive feature extraction through the combination of deformable convolution and transfer learning, cannot improve the classification accuracy under complex terrain, cannot quantify the spatiotemporal response of biological communities by constructing particle evolution models and simultaneously construct a multidimensional abiotic value indicator system to evaluate land degradation, and cannot achieve data transparency and support cross-institutional collaborative governance.

[0005] To solve the above technical problems, the present invention provides the following technical solution: a method for assessing land degradation status, comprising the following steps:

[0006] Step S1: multi-source remote sensing is used to collect land images and classify land organisms by extracting remote sensing spatial features at multiple levels;

[0007] Step S2: constructing a spatiotemporal coupled dynamic particle evolution model based on the land organism distribution map to predict the dynamic particle evolution of land organisms;

[0008] Step S3: Establish an abiotic value quantification model, calculate the historical value of the abiotic value of land, assign weights based on historical data and machine learning algorithms, and dynamically calculate a multi-dimensional value indicator system;

[0009] Step S4: Assess the land degradation status based on the historical value of land abiotic values and the dynamic particle evolution prediction of land organisms.

[0010] Preferably, remote sensing images of the target land are collected by multi-source remote sensing equipment, and the remote sensing images include multispectral images, thermal infrared images, lidar point cloud data, normalized vegetation index NDVI and Sentinel-1 SAR dual-polarization data. The multispectral images include more than 11 spectral bands, and the spectral bands include 3 near-infrared narrowband channels dedicated to chlorophyll fluorescence detection.

[0011] Preferably, geometric correction, radiometric correction and spatiotemporal alignment operations are performed on the remote sensing image, and a deformable convolution model is used to extract first remote sensing spatial features, wherein the first remote sensing spatial features include spectral features, texture features and structural features;

[0012] Performing cross-modal feature alignment on the first remote sensing spatial features through a transfer learning pre-training model to obtain a second remote sensing spatial feature that is time-space aligned;

[0013] ResNet is used to extract the low-level features of the second remote sensing spatial features, and the low-level features are secondary mined into global dependencies and high-level features through Transformer. The global dependencies and high-level features are classified through the random forest model to obtain the land biological cover type classification. The land biological cover type classification is plotted into a particle distribution structure diagram through statistical software, and different types of organisms are represented on the same land biological distribution map through different colors and shapes.

[0014] Preferably, a spatiotemporally coupled dynamic particle evolution model is constructed based on the land biomass distribution map, and big data is used to collect the number or area loss rate of the current land biomass, meteorological data, and soil moisture time series data. The meteorological data includes temperature, precipitation, and wind speed. Each biomass is mapped into a particle unit with an adaptive quality parameter. The meteorological data is passed through a random forest to obtain outliers, and the outliers are compared with time to obtain an anomaly index.

[0015] A spatiotemporal feature matrix is constructed based on the land biological distribution map, wherein the columns of the spatiotemporal feature matrix represent time series and the rows represent biological types. The unit feature matrix of the spatiotemporal feature matrix is solved by basic matrix operations, and the node data corresponding to the unit feature matrix is standardized to obtain a solution space probability matrix of the spatiotemporal feature matrix. The solution space probability matrix and the second remote sensing spatial feature are weightedly merged by a sliding window in the column direction to obtain a spatiotemporal feature probability distribution matrix. The spatiotemporal feature probability distribution matrix is aligned with the time resolution of multi-source data through a sliding window method to generate a standardized spatiotemporal dataset.

[0016] Preferably, each biome in the land biomass distribution map is mapped as a dynamic particle unit carrying a gene entropy value, an energy metabolic rate, and an environmental response function, wherein the gene entropy value is obtained by calculating the species genetic diversity index, and the environmental response function is obtained by integrating the multidimensional characteristics of the number or area loss rate of the current land biome, meteorological data, and soil moisture time series data;

[0017] The gene entropy value, energy metabolism rate and environmental response function are used as particle parameters, the variance of each particle parameter of the dynamic particle unit corresponding to each biological community is calculated, and the F statistic of each particle parameter is calculated based on the variance.

[0018] Preferably, the F statistics of each particle parameter corresponding to each dynamic particle unit and the standardized spatiotemporal dataset are subjected to spatiotemporal weighted fusion and input into an LSTM-Transformer hybrid neural network, wherein the LSTM layer is used to process meteorological data and soil moisture time series data to extract long-range dependency features, and the Transformer multi-head attention mechanism is used to fuse spatial distribution features through the multi-head attention mechanism to capture the spatial correlation of cross-regional biological migration, and residual connections are used to bridge the output data of the macro or micro model to generate the particle swarm evolution trajectory corresponding to the dynamic particle unit in the future time window;

[0019] The probability cloud map of the particle swarm evolution trajectory in the future time window is output, and the prediction reliability is double-verified using Brier score and KL divergence. When the prediction value is lower than 85%, the online update of the model based on transfer learning is triggered.

[0020] Preferably, step S3 includes:

[0021] The value indicators of the historical value of the abiotic value of the land include the monetary value created on the same piece of land, the ecological environment status and ecosystem stability and the protection area of specific biological environments;

[0022] Establish a quantitative model for non-biological value, obtain historical economic operation data of the target land through the blockchain distributed ledger, simultaneously obtain satellite remote sensing ecological index and ecosystem service value equivalent database, adopt a multi-objective optimization algorithm empowered by the attention mechanism, and dynamically calculate the multi-dimensional value indicator system.

[0023] Preferably, the multidimensional value indicator system includes economic value indicators, ecological stability indicators and biological protection intensity indicators. The economic value indicators include GDP output per unit area, land transfer transaction price and agricultural product market value. The ecological stability indicators include biodiversity index, soil and water conservation coefficient and carbon sequestration capacity. The biological protection intensity indicators include the coverage rate of endangered species habitats, the proportion of nature reserves and the area of ecological red line areas.

[0024] The weight distribution adopts the improved AdaBoost ensemble learning framework, and uses the SMOTE-Tomek joint sampling algorithm to deal with the category imbalance problem of the multi-dimensional value indicator system data. The principal component features with a cumulative contribution rate of ≥85% are extracted through principal component analysis (PCA).

[0025] Integrating a cost-sensitive learning mechanism into the improved AdaBoost ensemble learning framework, defining a misclassification cost matrix according to the improved AdaBoost ensemble learning framework incorporating the cost-sensitive learning mechanism, including constructing a multi-head cross-attention mechanism module and using the improved AdaBoost ensemble learning framework for weight distribution;

[0026] Building a multi-head cross attention mechanism module includes:

[0027] The nonlinear correlation between economic value indicators and ecological stability indicators is calculated through graph neural network (GNN), and the dynamic changes of time series data are captured through gated temporal attention unit (GTAU).

[0028] The weight distribution using the improved AdaBoost ensemble learning framework includes:

[0029] Using the AdaBoost ensemble learning framework to train multiple weak classifiers, the outputs of all the weak classifiers are weighted and summed to form a strong classifier. When the strong classifier is trained iteratively, the weights are adjusted based on the error rate of the strong classifier. The adjustment logic includes: training the next strong classifier, and when the output of the next strong classifier focuses on a data point misclassified by the previous strong classifier, outputting the current weight as the final weight distribution to obtain the weighted economic value index and ecological stability index;

[0030] The empowered economic value index and ecological stability index are input into the blockchain smart contract, and the inter-period discount calculation is performed in combination with the time value factor. The time value factor includes the discount rate and the inflation rate. The mathematical expression of inter-period discount is:

[0031] ;

[0032] in, is the discounted value at time t, is the value at the current moment, is the discount rate and t is the time.

[0033] Preferably, step S4 includes:

[0034] The particle swarm evolution trajectory and multi-dimensional value indicator system are used as sub-models for local training using differential privacy protection technology. The central server aggregates the degradation features of each region through an adaptive knowledge distillation algorithm to build a degradation assessment model including:

[0035] The modified universal soil loss equation is used as the physical loss function, and its mathematical expression is:

[0036] A=R×K×LS×C×P;

[0037] Among them, A is the average annual soil loss, R is the rainfall erosion factor, K is the soil erodibility factor, LS is the topography factor, C is the vegetation cover factor, and P is the soil and water conservation measures factor.

[0038] The Neural Radiation Field (NeRF) technology is used to model the land degradation process in a spatiotemporal continuous manner, and a continuous spatiotemporal remote sensing land degradation model is obtained.

[0039] A convolutional neural network (CNN) is used to extract features from the continuous spatiotemporal remote sensing land degradation model, a U-Net architecture containing 64 convolution kernels is constructed, a spatial probability prediction of degradation hotspot areas is obtained, and a risk level is output. The risk level includes a first risk level, a second risk level, and a third risk level. The first risk level includes a degradation probability value of the spatial probability prediction less than 0.2, the second risk level includes a degradation probability value of the spatial probability prediction between 0.2 and 0.5, and the third risk level includes a degradation probability value of the spatial probability prediction greater than 0.5.

[0040] The continuous spatiotemporal remote sensing land degradation model was deployed on the Google Earth Engine platform. Multi-temporal Landsat and Sentinel data stacks were used to extract the boundaries of land features through object-oriented segmentation technology, and a three-level classification of wetlands, woodlands, and cultivated land was achieved by combining decision tree classification rules.

[0041] A land degradation status assessment system, which is used to implement a land degradation status assessment method, comprising a biological classification module, a biological evolution module, a value evolution module and a degradation assessment module;

[0042] The biological classification module is used to collect land images through multi-source remote sensing and classify land organisms by extracting remote sensing spatial features at multiple levels;

[0043] The biological evolution module is used to construct a spatiotemporal coupled dynamic particle evolution model based on the land biological distribution map, and to predict the dynamic particle evolution of land organisms;

[0044] The value evolution module is used to establish an abiotic value quantification model, calculate the historical value of the abiotic value of land, assign weights based on historical data and machine learning algorithms, and dynamically calculate a multi-dimensional value indicator system;

[0045] The degradation assessment module is used to assess the land degradation status based on the historical value of the land non-living matter and the dynamic particle evolution prediction of the land organisms.

[0046] The beneficial effects of the present invention are as follows: adaptive feature extraction is performed through the combination of deformable convolution and transfer learning, classification accuracy is improved under complex terrain, land degradation is assessed by quantifying the spatiotemporal response of biological communities through the construction of a particle evolution model and simultaneously constructing a multidimensional abiotic value indicator system, blockchain is enhanced in credibility, distributed ledgers ensure data transparency, and cross-institutional collaborative governance is supported. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram of the basic flow of a land degradation status assessment method provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0049] Reference Figure 1 , as an embodiment of the present invention, provides a method for assessing land degradation status, comprising the following steps:

[0050] Step S1: multi-source remote sensing is used to collect land images and classify land organisms by extracting remote sensing spatial features at multiple levels;

[0051] Step S2: constructing a spatiotemporal coupled dynamic particle evolution model based on the land organism distribution map to predict the dynamic particle evolution of land organisms;

[0052] Step S3: Establish an abiotic value quantification model, calculate the historical value of the abiotic value of land, assign weights based on historical data and machine learning algorithms, and dynamically calculate a multi-dimensional value indicator system;

[0053] Step S4: Assess the land degradation status based on the historical value of land abiotic values and the dynamic particle evolution prediction of land organisms.

[0054] This solution innovatively integrates multi-source data fusion of deep learning, biological evolution modeling of dynamic particle systems, and cutting-edge ecological value quantification technology enabled by blockchain, forming a land degradation assessment system that is technologically advanced and practical.

[0055] Remote sensing images of the target land are collected through multi-source remote sensing equipment. The remote sensing images include multispectral images, thermal infrared images, lidar point cloud data, normalized difference vegetation index (NDVI) and Sentinel-1 SAR dual-polarization data. The multispectral images include more than 11 spectral bands, including 3 near-infrared narrowband channels dedicated to chlorophyll fluorescence detection.

[0056] Perform geometric correction, radiometric correction, and spatiotemporal alignment on remote sensing images, and use a deformable convolutional model to extract more representative first remote sensing spatial features. This can effectively improve model performance, especially when the image has non-rigid deformation or geometric changes. The first remote sensing spatial features include spectral features, texture features, and structural features.

[0057] The first remote sensing spatial feature is aligned across modalities using a transfer learning pre-trained model to obtain a second remote sensing spatial feature that is aligned in time and space.

[0058] ResNet is used to extract the low-level features of the second remote sensing spatial features, and the low-level features are secondary mined into global dependencies and high-level features through Transformer. In this way, rich local features can be extracted and global spatial information can be captured. The global dependencies and high-level features are classified through the random forest model to obtain the land biological cover type classification. The land biological cover type classification is plotted into a particle distribution structure diagram through statistical software, and different types of organisms are represented on the same land biological distribution map through different colors and shapes.

[0059] Colors include the three primary colors, and shapes include squares, triangles, and circles.

[0060] Deformable convolution is a convolution method that introduces an offset mechanism on the basis of traditional convolution, which enables the convolution operation to adapt to objects or image areas with large shape changes. In land biological classification, the targets in remote sensing images may have irregular shapes and different scales. Therefore, the use of deformable convolution can effectively enhance the robustness of the network and improve the ability to recognize complex land features.

[0061] A spatiotemporal coupled dynamic particle evolution model was constructed based on the land biomass distribution map. Big data was used to collect the current land biome population or area loss rate, meteorological data, and soil moisture time series data. Meteorological data included temperature, precipitation, and wind speed. Each biome was mapped to a particle unit with adaptive quality parameters. The meteorological data was used through a random forest to obtain outliers. These outliers were compared with time to obtain an anomaly index.

[0062] A spatiotemporal feature matrix is constructed based on the land biological distribution map. The columns of the spatiotemporal feature matrix represent time series and the rows represent biological types. The unit feature matrix of the spatiotemporal feature matrix is solved through basic matrix operations. The node data corresponding to the unit feature matrix is standardized to obtain the solution space probability matrix of the spatiotemporal feature matrix. The solution space probability matrix and the second remote sensing spatial feature are weightedly merged in the sliding window direction in the column direction to obtain the spatiotemporal feature probability distribution matrix. The spatiotemporal feature probability distribution matrix is aligned with the time resolution of multi-source data through the sliding window method to generate a standardized spatiotemporal dataset.

[0063] Each biome in the land biomass distribution map is mapped as a dynamic particle unit carrying gene entropy, energy metabolism rate, and environmental response function. The gene entropy is obtained by calculating the species genetic diversity index, and the environmental response function is obtained by integrating the multidimensional characteristics of the current land biome population or area loss rate, meteorological data, and soil moisture time series data.

[0064] The gene entropy value reflects the level of genetic diversity, the environmental response function represents the sensitivity to temperature and humidity fluctuations, and the energy metabolism rate is calculated based on NDVI and photosynthetically active radiation;

[0065] Taking gene entropy, energy metabolic rate and environmental response function as particle parameters, the variance of each particle parameter of the dynamic particle unit corresponding to each biological community is calculated. The F statistic of each particle parameter is calculated based on the variance, and the high-potential search area is identified through the gradient boosting decision tree to enhance the local convergence efficiency.

[0066] The F statistics of each particle parameter corresponding to each dynamic particle unit and the standardized spatiotemporal dataset are subjected to spatiotemporal weighted fusion and input into the LSTM-Transformer hybrid neural network. The LSTM layer is used to process meteorological data and soil moisture time series data to extract long-range dependency features. The Transformer multi-head attention mechanism is used to fuse spatial distribution features through the multi-head attention mechanism to capture the spatial correlation of cross-regional biological migration. Residual connections are used to bridge the output data of the macro or micro model to generate the particle swarm evolution trajectory corresponding to the dynamic particle unit in the future time window.

[0067] The probability cloud map of the particle swarm evolution trajectory in the future time window is output, and the prediction reliability is double-verified using Brier score and KL divergence. When the prediction value is lower than 85%, the online update of the model based on transfer learning is triggered.

[0068] Step S3 includes:

[0069] The value indicators of the historical value of the abiotic value of land include the monetary value created on the same land, the ecological environment status and ecosystem stability and the protection area of specific biological environments;

[0070] Establish a quantitative model for non-biological value, obtain historical economic operation data of the target land through the blockchain distributed ledger, simultaneously obtain satellite remote sensing ecological index and ecosystem service value equivalent database, adopt a multi-objective optimization algorithm empowered by the attention mechanism, and dynamically calculate the multi-dimensional value indicator system.

[0071] The multidimensional value indicator system includes economic value indicators, ecological stability indicators, and biological protection intensity indicators. Economic value indicators include GDP output per unit area, land transfer transaction price, and agricultural product market value. Ecological stability indicators include biodiversity index, soil and water conservation coefficient, and carbon sequestration capacity. Biological protection intensity indicators include the habitat coverage rate of endangered species, the proportion of nature reserves, and the area of ecological red line areas.

[0072] The weight distribution adopts the improved AdaBoost ensemble learning framework, and uses the SMOTE-Tomek joint sampling algorithm to deal with the category imbalance problem of the multi-dimensional value indicator system data. The principal component features with a cumulative contribution rate of ≥85% are extracted through principal component analysis (PCA).

[0073] The improved AdaBoost ensemble learning framework is integrated with a cost-sensitive learning mechanism, and the misclassification cost matrix is defined based on the improved AdaBoost ensemble learning framework with the cost-sensitive learning mechanism, including the construction of a multi-head cross-attention mechanism module and the use of the improved AdaBoost ensemble learning framework for weight distribution.

[0074] Building a multi-head cross attention mechanism module includes:

[0075] The nonlinear correlation between economic value indicators and ecological stability indicators is calculated through the graph neural network (GNN), and the dynamic changes of time series data are captured through the gated temporal attention unit (GTAU). The gating mechanism can adjust the contribution of each time step by learning the weight of the sequence, and the self-attention mechanism is used to capture the lag effect and mutation characteristics.

[0076] The weight distribution using the improved AdaBoost ensemble learning framework includes:

[0077] The AdaBoost ensemble learning framework is used to train multiple weak classifiers. The weighted sum of all weak classifier outputs is used to form a strong classifier. When the strong classifier is trained iteratively, the weights are adjusted based on the error rate of the strong classifier. The adjustment logic includes: training the next strong classifier. When the output of the next strong classifier focuses on a data point misclassified by the previous strong classifier, the current weight is output as the final weight distribution, resulting in the weighted economic value index and ecological stability index.

[0078] The empowered economic value index and ecological stability index are input into the blockchain smart contract, and the inter-period discount calculation is performed in combination with the time value factor. The time value factor includes the discount rate and the inflation rate. The discount rate at each time point is taken into account to estimate the future economic and ecological value. The inter-period discount mathematical expression is:

[0079] ;

[0080] in, is the discounted value at time t, is the value at the current moment, is the discount rate and t is the time.

[0081] Step S4 includes:

[0082] The particle swarm evolution trajectory and multi-dimensional value indicator system are used as sub-models for local training using differential privacy protection technology. The central server aggregates the degradation characteristics of each region through an adaptive knowledge distillation algorithm to build a degradation assessment model including:

[0083] The modified universal soil loss equation is used as the physical loss function, and its mathematical expression is:

[0084] A=R×K×LS×C×P;

[0085] Among them, A is the average annual soil loss, R is the rainfall erosion factor, K is the soil erodibility factor, LS is the topography factor, C is the vegetation cover factor, and P is the soil and water conservation measures factor.

[0086] The Neural Radiation Field (NeRF) technology is used to model the land degradation process in a spatiotemporal continuous manner, and a continuous spatiotemporal remote sensing land degradation model is obtained.

[0087] A convolutional neural network (CNN) was used to extract features from the continuous spatiotemporal remote sensing land degradation model. A U-Net architecture containing 64 convolution kernels was constructed to obtain spatial probability predictions of degradation hotspots and output risk levels. The risk levels included first, second, and third risk levels. The first risk level included degradation probability values less than 0.2 from the spatial probability prediction, the second risk level included degradation probability values between 0.2 and 0.5 from the spatial probability prediction, and the third risk level included degradation probability values greater than 0.5 from the spatial probability prediction.

[0088] The continuous spatiotemporal remote sensing land degradation model was deployed on the Google Earth Engine platform. Multi-temporal Landsat and Sentinel data stacks were used to extract the boundaries of land features through object-oriented segmentation technology, and a three-level classification of wetlands, woodlands, and cultivated land was achieved by combining decision tree classification rules.

[0089] A land degradation status assessment system, which is used to implement a land degradation status assessment method, comprising a biological classification module, a biological evolution module, a value evolution module and a degradation assessment module;

[0090] The biological classification module is used to collect land images from multi-source remote sensing and classify land organisms by extracting remote sensing spatial features at multiple levels;

[0091] The biological evolution module is used to build a spatiotemporal coupled dynamic particle evolution model based on the land biological distribution map and predict the dynamic particle evolution of land organisms;

[0092] The value evolution module is used to establish a quantitative model for abiotic value, calculate the historical value of abiotic land, assign weights based on historical data and machine learning algorithms, and dynamically calculate a multi-dimensional value indicator system;

[0093] The degradation assessment module is used to assess the land degradation status based on the historical value of land abiotic values and the dynamic particle evolution prediction of land organisms.

[0094] This invention combines deformable convolution with transfer learning to perform adaptive feature extraction, improving classification accuracy in complex terrain. It evaluates land degradation by constructing a particle evolution model to quantify the spatiotemporal response of biological communities and simultaneously constructing a multidimensional abiotic value indicator system. This enables blockchain to enhance credibility, distributed ledgers to ensure data transparency, and support cross-institutional collaborative governance.

[0095] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium may be implemented by any type of volatile or non-volatile storage device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for assessing land degradation, characterized in that: The following steps are involved: Step S1: multi-source remote sensing is used to collect land images and classify land organisms by extracting remote sensing spatial features at multiple levels; Step S2: constructing a spatiotemporal coupled dynamic particle evolution model based on the land organism distribution map to predict the dynamic particle evolution of land organisms; Step S3: Establish an abiotic value quantification model, calculate the historical value of the abiotic value of land, assign weights based on historical data and machine learning algorithms, and dynamically calculate a multi-dimensional value indicator system; Step S4: assessing the land degradation status based on the historical values of land abiotic values and the dynamic particle evolution prediction of land biotic values; A spatiotemporally coupled dynamic particle evolution model is constructed based on the land biomass distribution map. Big data is used to collect the current land biome population or area loss rate, meteorological data, and soil moisture time series data. The meteorological data includes temperature, precipitation, and wind speed. Each biome is mapped into a particle unit with an adaptive quality parameter. The meteorological data is passed through a random forest to obtain outliers. The outliers are compared with time to obtain an anomaly index. A spatiotemporal feature matrix is constructed based on a land biological distribution map, wherein the columns of the spatiotemporal feature matrix represent time series and the rows represent biological types; a unit feature matrix of the spatiotemporal feature matrix is obtained by matrix basic operations; the node data corresponding to the unit feature matrix is standardized to obtain a solution space probability matrix of the spatiotemporal feature matrix; the solution space probability matrix is weightedly merged with the second remote sensing spatial feature using a sliding window in the column direction to obtain a spatiotemporal feature probability distribution matrix; the spatiotemporal feature probability distribution matrix is aligned with the time resolution of multi-source data using a sliding window method to generate a standardized spatiotemporal dataset; Each biome in the land biomass distribution map is mapped as a dynamic particle unit carrying gene entropy, energy metabolism rate, and environmental response function. The gene entropy is obtained by calculating the species genetic diversity index, and the environmental response function is obtained by integrating the multidimensional characteristics of the current land biome population or area loss rate, meteorological data, and soil moisture time series data. The gene entropy value, energy metabolic rate and environmental response function are used as particle parameters, the variance of each particle parameter of the dynamic particle unit corresponding to each biological community is calculated, and the F statistic of each particle parameter is calculated based on the variance; The F statistics of each particle parameter corresponding to each dynamic particle unit and the standardized spatiotemporal dataset are subjected to spatiotemporal weighted fusion and input into the LSTM-Transformer hybrid neural network. The LSTM layer is used to process meteorological data and soil moisture time series data to extract long-range dependency features. The Transformer multi-head attention mechanism is used to fuse spatial distribution features through the multi-head attention mechanism to capture the spatial correlation of cross-regional biological migration. Residual connections are used to bridge the output data of the macro or micro model to generate the particle swarm evolution trajectory corresponding to the dynamic particle unit in the future time window. The probability cloud map of the particle swarm evolution trajectory in the future time window is output, and the prediction reliability is double-verified using Brier score and KL divergence. When the prediction value is lower than 85%, the online update of the model based on transfer learning is triggered.

2. The land degradation status assessment method according to claim 1, wherein: Remote sensing images of the target land are collected using multi-source remote sensing equipment. The remote sensing images include multispectral images, thermal infrared images, lidar point cloud data, normalized difference vegetation index (NDVI), and Sentinel-1 SAR dual-polarization data. The multispectral images include more than 11 spectral bands, including three near-infrared narrowband channels dedicated to chlorophyll fluorescence detection.

3. The land degradation assessment method according to claim 2, wherein: Performing geometric correction, radiometric correction, and spatiotemporal alignment operations on the remote sensing image, and extracting first remote sensing spatial features using a deformable convolution model, wherein the first remote sensing spatial features include spectral features, texture features, and structural features; Performing cross-modal feature alignment on the first remote sensing spatial features through a transfer learning pre-training model to obtain a second remote sensing spatial feature that is time-space aligned; ResNet is used to extract the low-level features of the second remote sensing spatial features, and the low-level features are secondary mined into global dependencies and high-level features through Transformer. The global dependencies and high-level features are classified through the random forest model to obtain the land biological cover type classification. The land biological cover type classification is plotted into a particle distribution structure diagram through statistical software, and different types of organisms are represented on the same land biological distribution map through different colors and shapes.

4. The land degradation status assessment method according to claim 3, wherein: The step S3 comprises: The value indicators of the historical value of the abiotic value of the land include the monetary value created on the same piece of land, the ecological environment status and ecosystem stability and the protection area of specific biological environments; Establish a quantitative model for non-biological value, obtain historical economic operation data of the target land through the blockchain distributed ledger, simultaneously obtain satellite remote sensing ecological index and ecosystem service value equivalent database, adopt a multi-objective optimization algorithm empowered by the attention mechanism, and dynamically calculate the multi-dimensional value indicator system.

5. The land degradation assessment method according to claim 4, wherein: The multi-dimensional value indicator system includes economic value indicators, ecological stability indicators, and biological protection intensity indicators. The economic value indicators include GDP output per unit area, land transfer transaction price, and agricultural product market value. The ecological stability indicators include biodiversity index, soil and water conservation coefficient, and carbon sequestration capacity. The biological protection intensity indicators include the coverage rate of endangered species habitats, the proportion of nature reserves, and the area of ecological red line areas. The weight distribution adopts the improved AdaBoost ensemble learning framework, and uses the SMOTE-Tomek joint sampling algorithm to deal with the category imbalance problem of the multi-dimensional value indicator system data. The principal component features with a cumulative contribution rate of ≥85% are extracted through principal component analysis (PCA). Integrating a cost-sensitive learning mechanism into the improved AdaBoost ensemble learning framework, defining a misclassification cost matrix according to the improved AdaBoost ensemble learning framework incorporating the cost-sensitive learning mechanism, including constructing a multi-head cross-attention mechanism module and using the improved AdaBoost ensemble learning framework for weight distribution; Building a multi-head cross attention mechanism module includes: The nonlinear correlation between economic value indicators and ecological stability indicators is calculated through the graph neural network (GNN), and the dynamic changes of time series data are captured through the gated temporal attention unit (GTAU). The weight distribution using the improved AdaBoost ensemble learning framework includes: Using the AdaBoost ensemble learning framework to train multiple weak classifiers, the outputs of all the weak classifiers are weighted and summed to form a strong classifier. When the strong classifier is trained iteratively, the weights are adjusted based on the error rate of the strong classifier. The adjustment logic includes: training the next strong classifier, and when the output of the next strong classifier focuses on a data point misclassified by the previous strong classifier, outputting the current weight as the final weight distribution to obtain the weighted economic value index and ecological stability index; The empowered economic value index and ecological stability index are input into the blockchain smart contract, and the inter-period discount calculation is performed in combination with the time value factor. The time value factor includes the discount rate and the inflation rate. The mathematical expression of inter-period discount is: ； in, is the discounted value at time t, is the value at the current moment, is the discount rate and t is the time.

6. The land degradation status assessment method according to claim 5, characterized in that: The step S4 comprises: The particle swarm evolution trajectory and multi-dimensional value indicator system are used as sub-models for local training using differential privacy protection technology. The central server aggregates the degradation features of each region through an adaptive knowledge distillation algorithm to build a degradation assessment model including: The modified universal soil loss equation is used as the physical loss function, and its mathematical expression is: A=R×K×LS×C×P; Among them, A is the average annual soil loss, R is the rainfall erosion factor, K is the soil erodibility factor, LS is the topography factor, C is the vegetation cover factor, and P is the soil and water conservation measures factor; The Neural Radiation Field (NeRF) technology is used to model the land degradation process in a spatiotemporal continuous manner, and a continuous spatiotemporal remote sensing land degradation model is obtained. A convolutional neural network (CNN) was used to extract features from the continuous spatiotemporal remote sensing land degradation model, and a U-Net architecture containing 64 convolution kernels was constructed to obtain spatial probability predictions of degradation hotspot areas and output risk levels. The risk levels included a first risk level, a second risk level, and a third risk level. The first risk level included a degradation probability value of less than 0.2 in the spatial probability prediction, a second risk level included a degradation probability value between 0.2 and 0.5 in the spatial probability prediction, and a third risk level included a degradation probability value greater than 0.5 in the spatial probability prediction: The continuous spatiotemporal remote sensing land degradation model was deployed on the Google Earth Engine platform. Multi-temporal Landsat and Sentinel data stacks were used to extract the boundaries of land features through object-oriented segmentation technology, and a three-level classification of wetlands, woodlands, and cultivated land was achieved by combining decision tree classification rules.

7. A land degradation status assessment system, the system being used to implement a land degradation status assessment method, characterized in that: Including biological classification module, biological evolution module, value evolution module and degradation assessment module: The biological classification module is used to collect land images through multi-source remote sensing and classify land organisms by extracting remote sensing spatial features at multiple levels; The biological evolution module is used to construct a spatiotemporal coupled dynamic particle evolution model based on the land biological distribution map, and to predict the dynamic particle evolution of land organisms; The value evolution module is used to establish an abiotic value quantification model, calculate the historical value of the abiotic value of land, assign weights based on historical data and machine learning algorithms, and dynamically calculate a multi-dimensional value indicator system; The degradation assessment module is used to assess the land degradation status based on the historical value of land abiotic values and the dynamic particle evolution prediction of land organisms; A spatiotemporally coupled dynamic particle evolution model is constructed based on the land biomass distribution map. Big data is used to collect the current land biome population or area loss rate, meteorological data, and soil moisture time series data. The meteorological data includes temperature, precipitation, and wind speed. Each biome is mapped into a particle unit with an adaptive quality parameter. The meteorological data is passed through a random forest to obtain outliers. The outliers are compared with time to obtain an anomaly index. A spatiotemporal feature matrix is constructed based on a land biological distribution map, wherein the columns of the spatiotemporal feature matrix represent time series and the rows represent biological types; a unit feature matrix of the spatiotemporal feature matrix is obtained by matrix basic operations; the node data corresponding to the unit feature matrix is standardized to obtain a solution space probability matrix of the spatiotemporal feature matrix; the solution space probability matrix is weightedly merged with the second remote sensing spatial feature using a sliding window in the column direction to obtain a spatiotemporal feature probability distribution matrix; the spatiotemporal feature probability distribution matrix is aligned with the time resolution of multi-source data using a sliding window method to generate a standardized spatiotemporal dataset; Each biome in the land biomass distribution map is mapped as a dynamic particle unit carrying gene entropy, energy metabolism rate, and environmental response function. The gene entropy is obtained by calculating the species genetic diversity index, and the environmental response function is obtained by integrating the multidimensional characteristics of the current land biome population or area loss rate, meteorological data, and soil moisture time series data. The gene entropy value, energy metabolic rate and environmental response function are used as particle parameters, the variance of each particle parameter of the dynamic particle unit corresponding to each biological community is calculated, and the F statistic of each particle parameter is calculated based on the variance; The F statistics of each particle parameter corresponding to each dynamic particle unit and the standardized spatiotemporal dataset are subjected to spatiotemporal weighted fusion and input into the LSTM-Transformer hybrid neural network. The LSTM layer is used to process meteorological data and soil moisture time series data to extract long-range dependency features. The Transformer multi-head attention mechanism is used to fuse spatial distribution features through the multi-head attention mechanism to capture the spatial correlation of cross-regional biological migration. Residual connections are used to bridge the output data of the macro or micro model to generate the particle swarm evolution trajectory corresponding to the dynamic particle unit in the future time window. The probability cloud map of the particle swarm evolution trajectory in the future time window is output, and the prediction reliability is double-verified using Brier score and KL divergence. When the prediction value is lower than 85%, the online update of the model based on transfer learning is triggered.

Images