A method and device for monitoring soil erosion by fusing multi-source remote sensing data

Abstract

The application provides a soil and water loss monitoring method and device for multi-source remote sensing data fusion, and relates to the technical field of deep learning. The method comprises the following steps: based on the collected optical remote sensing data and synthetic aperture radar data for monitoring landform areas, a cross-modal physical response consistency mapping relationship is established, and consistency correction processing is performed; after processing, shared semantic features are extracted from different remote sensing data, and structure compensation and sensitive weight adjustment are performed; in the process of extracting shared semantic features, the features of the confused area in the optical remote sensing data are corrected; the confidence area of different remote sensing features in the fusion feature space is processed, and the fusion features are taken as input to drive the soil and water loss spatial partition extraction and the soil and water loss quantitative evaluation, so that the soil and water loss monitoring is realized. The application establishes a cross-modal physical response consistency modeling and semantic shared feature optimization mechanism for multi-source remote sensing data, and improves the physical rationality and semantic expression ability of the fusion result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and in particular to a water and soil loss monitoring method and device for multi-source remote sensing data fusion. BACKGROUND

[0002] Water and soil loss monitoring refers to a technical process of dynamically observing, quantitatively analyzing and spatially evaluating the processes of erosion, transport and deposition of surface soil under the action of rainfall, runoff, wind or human disturbance through systematic means, aiming to identify the spatial range, evolution trend and intensity change of water and soil loss, and to evaluate the ecological environment risk and land degradation degree of the watershed. Water and soil loss monitoring usually combines ground measurement investigation and remote sensing data acquisition, and establishes a monitoring system of multi-temporal, multi-scale and multi-source data fusion through topographic analysis, surface cover change detection, soil erosion model driving, meteorological element correlation reasoning and other ways, to realize continuous tracking and accurate early warning of water and soil loss dynamics at complex landform area, small watershed and even regional scale, and to provide scientific basis for water and soil conservation engineering layout, ecological restoration strategy formulation and comprehensive management of the watershed.

[0003] Multi-source remote sensing monitoring of water and soil loss usually includes optical remote sensing measurement and radar remote sensing measurement. Optical remote sensing measurement of water and soil loss mainly relies on high-resolution capture of visible light, near-infrared and short-wave infrared band radiation signals reflected by the ground, and extracts the spatial distribution and change dynamics of water and soil loss patches, erosion gullies and debris flow areas by analyzing the ground vegetation cover index, bare ground index, soil color change and ground humidity characteristics, with advantages of rich wavebands, fine classification, high spatio-temporal resolution, etc., but it is easily disturbed by observation under cloudy, rainy or shielding conditions. Radar remote sensing measurement uses synthetic aperture radar system to actively emit microwave signals and receive their ground scattering echoes, and based on backscattering intensity, polarization characteristics and interference change analysis, it inverts ground roughness, soil moisture and micro-topographic disturbance process, and can penetrate clouds and part of vegetation cover, realizing all-weather, all-time water and soil loss monitoring, especially suitable for extracting bare surface erosion degree, landslide volume change and fine-scale landform deformation. The collaborative fusion of optical remote sensing measurement and radar remote sensing measurement has become an important trend to improve the accuracy and reliability of dynamic monitoring of water and soil loss.

[0004] The current multi-source remote sensing data fusion algorithm generally adopts a surface feature combination method such as pixel-level superposition, feature splicing or decision-level voting, and usually only relies on experiential normalization to simply process different source data, lacks systematic modeling and physical consistency alignment of deep essential differences such as observation mechanism, response scale, physical quantity level and error characteristics between optical remote sensing data and synthetic aperture radar remote sensing data, and causes information redundancy, modal conflict amplification and feature class recognition error between different modal features in the fusion process, which seriously affects the reliability of the fusion features and the accuracy and stability of the downstream soil and water loss monitoring task, and therefore it is urgent to establish a cross-modal physical response consistency modeling and semantic shared feature optimization mechanism for multi-source remote sensing data to improve the physical rationality and semantic expression ability of the fusion results. SUMMARY

[0005] The present application provides a soil and water loss monitoring method and device for multi-source remote sensing data fusion, establishes a cross-modal physical response consistency modeling and semantic shared feature optimization mechanism for multi-source remote sensing data, and improves the physical rationality and semantic expression ability of the fusion results.

[0006] In a first aspect of the present application, a soil and water loss monitoring method for multi-source remote sensing data fusion is provided, which comprises:

[0007] Based on the collected optical remote sensing data and synthetic aperture radar data for the monitored landform area, a cross-modal physical response consistency mapping relationship is established, and a physical scale standardization mechanism is used to perform consistency correction processing on the physical quantity level, response scale and radiation dynamic range of the optical remote sensing data and the synthetic aperture radar remote sensing data;

[0008] After completing the physical scale standardization processing, shared semantic features are extracted from the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data, and a multi-scale residual structure and a dynamic spatial attention mechanism are used to perform structure compensation and sensitive weight adjustment on the optical remote sensing features and the synthetic aperture radar remote sensing features, wherein the optical remote sensing features are features corresponding to the corrected optical remote sensing data, and the synthetic aperture radar remote sensing features are features corresponding to the corrected synthetic aperture radar remote sensing data;

[0009] In the process of extracting shared semantic features, the structural texture features of the synthetic aperture radar remote sensing data are used to correct the confused area features in the optical remote sensing data by guiding learning a distillation network, so as to suppress modal conflict features and improve boundary recognition ability;

[0010] After the fusion features are generated according to the processed optical remote sensing features and the processed synthetic aperture radar remote sensing features, confidence regions in the fusion feature space are removed and weighted, and the fusion features are taken as inputs to drive the water and soil loss spatial partition extraction and the water and soil loss quantitative evaluation, so that the water and soil loss distribution pattern and the spatiotemporal evolution trend of the monitored landform region are monitored and evaluated.

[0011] On the basis of the above technical solutions, preferably, the cross-modal physical response consistency mapping relationship is established based on the collected optical remote sensing data and synthetic aperture radar data for the monitored landform region, and specifically includes:

[0012] Based on the ground object type division result of the monitored landform region, a ground object spectral reflectance characteristic model of the optical remote sensing data and a ground object electromagnetic scattering response model of the synthetic aperture radar remote sensing data are respectively constructed to clarify the physical response mechanism of different ground object categories under different remote sensing data;

[0013] Based on the ground object spectral reflectance characteristic model and the ground object electromagnetic scattering response model, a mathematical mapping relationship of the optical remote sensing data and the synthetic aperture radar remote sensing data in the physical quantity level, the response scale and the radiation dynamic range is established, so that the observation responses of the optical remote sensing data and the synthetic aperture radar remote sensing data to the same ground object category have alignment.

[0014] On the basis of the above technical solutions, preferably, the optical remote sensing data and the synthetic aperture radar remote sensing data are subjected to consistency correction processing of the physical quantity level, the response scale and the radiation dynamic range through a unified physical scale standardization mechanism, and specifically includes:

[0015] According to the mathematical mapping relationship, a unified physical scale standardization mechanism is designed to normalize the optical remote sensing data and the synthetic aperture radar remote sensing data to a unified physical dimension and dynamic range, so as to eliminate the observation scale deviation caused by the difference in sensor characteristics;

[0016] Based on the normalized optical remote sensing data and the normalized synthetic aperture radar remote sensing data, the physical consistency of each corresponding ground surface unit of the monitored landform region is verified to ensure that the consistency of the physical response meets the preset threshold requirement, and the data regions that do not meet the consistency requirement are subjected to mask processing, and the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data are obtained after the processing.

[0017] Based on the above technical solutions, preferably, after completing the physical scale standardization processing, the corrected optical remote sensing data and the synthetic aperture radar remote sensing data are extracted to share semantic features, a multi-scale residual structure and a dynamic spatial attention mechanism are used to compensate the structure and adjust the sensitive weight of the optical remote sensing features and the synthetic aperture radar remote sensing features, specifically including:

[0018] The corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data are respectively subjected to multi-scale feature extraction, and different scale size convolution kernels are used to extract the optical remote sensing features and the synthetic aperture radar remote sensing features;

[0019] In the process of extracting the optical remote sensing features and the synthetic aperture radar remote sensing features, a multi-scale residual structure is superimposed, and a cross-scale residual connection method is used to retain key boundary information and structural detail information in each scale feature;

[0020] Based on the optical remote sensing features and the synthetic aperture radar remote sensing features output by the multi-scale residual structure, a dynamic spatial attention mechanism is introduced, and position-sensitive weight maps are generated for optical remote sensing feature channels and synthetic aperture radar remote sensing feature channels respectively, and the importance of features at different spatial positions is dynamically weighted and modulated;

[0021] Based on the weighted results of the dynamic spatial attention mechanism, the optical remote sensing features and the synthetic aperture radar remote sensing features are enhanced in sensitive areas and suppressed in redundant areas;

[0022] The optical remote sensing features and the synthetic aperture radar remote sensing features subjected to multi-scale residual structure compensation and dynamic spatial attention mechanism weighting adjustment are used as the shared semantic features.

[0023] Based on the above technical solutions, preferably, in the process of extracting shared semantic features, a guided learning distillation network is used to correct the features of the confusion area in the optical remote sensing data based on the structural texture features of the synthetic aperture radar remote sensing data, suppress the modal conflict features and improve the boundary recognition ability, specifically including:

[0024] Based on the corrected synthetic aperture radar remote sensing data, a structural texture feature extraction is used to extract the ground fine-grained texture pattern to generate a synthetic aperture radar structural texture feature map;

[0025] Based on the corrected optical remote sensing data, an optical remote sensing confusion area feature map is extracted, wherein the confusion area is a spatial position with fuzzy boundary or overlapping ground object category;

[0026] The guided learning distillation network is constructed, the synthetic aperture radar structure texture feature map is taken as a guided signal, the optical remote sensing confusion area feature map is taken as a learning signal, a structure preservation loss function and a modal consistency loss function are combined for optimization, and the optical remote sensing confusion area feature map is driven to gradually approach the spatial distribution mode of the synthetic aperture radar structure texture feature map;

[0027] In the guided learning distillation process, the boundary position of the confusion area is corrected through local alignment in the distillation network, and modal conflict features and artifact features in the optical remote sensing confusion area feature map are suppressed, so that the spatial boundary definition and semantic distinction are improved.

[0028] The optical remote sensing feature is corrected by the guided learning distillation network to obtain the processed optical remote sensing feature, and the synthetic aperture radar remote sensing feature is corrected by the guided learning distillation network to obtain the processed synthetic aperture radar remote sensing feature.

[0029] On the basis of the above technical solutions, preferably, after the fusion feature is generated according to the processed optical remote sensing feature and the processed synthetic aperture radar remote sensing feature, the confidence area in the fusion feature space is subjected to elimination processing and confidence weighting processing, and the fusion feature is taken as input to drive water and soil loss spatial partition extraction and water and soil loss quantitative evaluation, so as to realize monitoring and evaluation of the water and soil loss distribution pattern and the spatio-temporal evolution trend of the monitored landform area, specifically including:

[0030] The processed optical remote sensing feature and the processed synthetic aperture radar remote sensing feature are taken as input to generate the fusion feature through a feature concatenation and a cross-modal attention fusion mechanism, and the fusion feature contains optical remote sensing feature information and synthetic aperture radar remote sensing feature information.

[0031] The fusion feature is subjected to confidence prediction based on a Bayesian deep neural network to generate a confidence map corresponding to each spatial position in the fusion feature space, and the confidence map represents the confidence level of the fusion feature at each position.

[0032] According to the spatial distribution result of the confidence map, a confidence elimination threshold is set, and a low confidence area in the confidence map that is lower than the elimination threshold is subjected to mask processing or zero processing in the fusion feature.

[0033] For a high confidence area in the confidence map that is higher than a confidence weighting threshold, a feature response is subjected to weighted amplification processing according to the confidence value, to obtain a feature fusion area.

[0034] The fusion features of the feature fusion area are input into a region connectivity analysis and change detection algorithm to extract soil erosion patch distribution patterns, and soil erosion intensity indexes are inverted to realize identification of soil erosion distribution patterns in the geomorphic area and dynamic monitoring of the spatiotemporal evolution trend.

[0035] On the basis of the above technical solutions, preferably, after the physical scale standardization processing is completed, the shared semantic features are extracted from the corrected optical remote sensing data and the synthetic aperture radar remote sensing data, and specifically, the method further comprises the following steps:

[0036] The optical remote sensing features and the synthetic aperture radar remote sensing features processed by the multi-scale residual structure and the dynamic spatial attention mechanism are mapped to the same shared semantic feature space to obtain shared optical remote sensing features and shared radar remote sensing features;

[0037] In the shared semantic feature space, a standardized class center vector is defined for each ground object class, a semantic consistency loss function is constructed to constrain the shared semantic features of the same class samples to be close to the same class center vector and the shared semantic features of different class samples to be far away from other class centers;

[0038] The mutual information amount of the shared optical remote sensing features and the shared radar remote sensing features in the shared semantic feature space is calculated, and a cross-modal mutual information maximization mechanism is adopted to maximize the mutual information between the shared optical remote sensing features and the shared radar remote sensing features;

[0039] In the process of calculating the mutual information amount, the semantic consistency loss function and the cross-modal mutual information maximization loss function are jointly optimized, and the balance between the class distinguishability and the modal reservation of the shared semantic features is maintained, and the shared optical remote sensing features and the shared radar remote sensing features are optimized respectively;

[0040] The optimized shared optical remote sensing features are used as the guide signal, and the optimized shared radar remote sensing features are used as the learning signal.

[0041] In a second aspect of the present application, a multi-source remote sensing data fusion soil erosion monitoring device is provided, which is used to perform any one of the multi-source remote sensing data fusion soil erosion monitoring methods described above, and the device comprises an acquisition module, a processing module and an output module, wherein:

[0042] The acquisition module is used to establish a cross-modal physical response consistency mapping relationship based on the collected optical remote sensing data and synthetic aperture radar data for the monitored geomorphic area, and to perform consistency correction processing on the physical quantity level, response scale and radiation dynamic range of the optical remote sensing data and the synthetic aperture radar remote sensing data through a unified physical scale standardization mechanism;

[0043] The processing module is configured to, after completing the physical scale normalization processing, extract shared semantic features from the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data, and use a multi-scale residual structure and a dynamic spatial attention mechanism to compensate the structure and adjust the sensitive weight of optical remote sensing features and synthetic aperture radar remote sensing features, wherein the optical remote sensing features are features corresponding to the corrected optical remote sensing data, and the synthetic aperture radar remote sensing features are features corresponding to the corrected synthetic aperture radar remote sensing data.

[0044] The processing module is configured to, in the process of extracting shared semantic features, correct the feature of the confusion area in the optical remote sensing data by guiding the learning of the structure texture feature of the synthetic aperture radar remote sensing data, suppress the modal conflict feature, and improve the boundary recognition ability.

[0045] The output module is configured to, after generating the fusion features from the processed optical remote sensing features and the processed synthetic aperture radar remote sensing features, perform rejection processing and confidence weighting processing on the confidence areas in the fusion feature space, and drive the water and soil loss spatial partition extraction and the water and soil loss quantitative evaluation by taking the fusion features as input, so as to realize the monitoring and evaluation of the water and soil loss distribution pattern and the spatio-temporal evolution trend of the monitored landform area.

[0046] In a third aspect of the present application, an electronic device is provided, which includes a processor, a memory, a user interface, and a network interface, the memory is configured to store instructions, the user interface and the network interface are configured to communicate with other devices, and the processor is configured to execute the instructions stored in the memory to enable the electronic device to perform the method according to any one of the preceding aspects.

[0047] In a fourth aspect of the present application, a computer readable storage medium is provided, which stores instructions, and when the instructions are executed, the method according to any one of the preceding aspects is performed.

[0048] In summary, the one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0049] 1. The present application introduces a cross-modal physical response consistency modeling mechanism in the process of multi-source remote sensing data fusion. First, based on the division of ground object types in the monitored landform area, ground object spectral reflectance models for optical remote sensing data and ground object electromagnetic scattering response models for synthetic aperture radar remote sensing data are constructed respectively. A unified mathematical mapping relationship and physical scale standardization mechanism are established to ensure the consistency of optical remote sensing data and synthetic aperture radar remote sensing data in physical magnitude, response scale, and radiation dynamic range. On this basis, multi-scale residual structure and dynamic spatial attention mechanism are used to extract shared semantic features after structure compensation. Guided learning distillation network is used to correct optical remote sensing confusion area features with synthetic aperture radar structural texture features. Semantic consistency loss function and cross-modal mutual information maximization mechanism are further introduced to jointly optimize the shared semantic feature space distribution, thereby ensuring the consistency of the fused data at the physical level and strengthening the collaborative expression of different modal features at the semantic level. Finally, the physical rationality and semantic expression ability of the fused features are significantly improved, and high-precision, high-robustness, and high-spatiotemporal resolution dynamic evolution analysis in soil and water loss monitoring tasks are achieved.

[0050] 2. Based on the division of ground object types, ground object spectral reflectance models for optical remote sensing data and ground object electromagnetic scattering response models for synthetic aperture radar remote sensing data are constructed respectively, and a mathematical mapping relationship is established to realize the unified alignment of optical remote sensing data and synthetic aperture radar remote sensing data in physical magnitude, response scale, and radiation dynamic range, thereby ensuring the consistency of different remote sensing data sources for observing the same ground object category and improving the physical rationality of subsequent fusion modeling and the interoperability of cross-modal data.

[0051] 3. The optical remote sensing data and synthetic aperture radar remote sensing data are normalized to a unified physical scale through a unified physical scale standardization mechanism, and after standardization, physical consistency verification and mask processing are performed on each ground surface unit to eliminate scale deviation and noise interference caused by differences in observation characteristics of different sensors, ensuring that the input feature data has high consistency and high reliability, providing stable and accurate basic data support for subsequent shared semantic feature extraction and fusion feature generation.

[0052] 4. In the feature extraction process, multi-scale residual structure and dynamic spatial attention mechanism are introduced to perform multi-scale structure compensation and spatial sensitive weight adjustment on optical remote sensing features and synthetic aperture radar remote sensing features, effectively enhancing the expression ability of ground object boundaries, change areas, and fine-grained texture features, while suppressing redundant irrelevant area features, significantly improving the discrimination ability and spatial expression accuracy of shared semantic features.

[0053] 5. By guiding learning distillation network, the optical remote sensing confusion area feature is corrected with synthetic aperture radar structural texture feature as a guide signal, and the modal conflict feature and artifact feature are suppressed in the local alignment process, which effectively improves the discrimination accuracy and spatial consistency of the boundary fuzzy area in the optical remote sensing data, thereby improving the expressiveness and stability of the shared semantic feature in change detection and small-scale ground disturbance identification.

[0054] 6. After the fusion feature is generated, the fusion feature confidence is predicted by a Bayesian deep neural network, and the low-confidence area is removed and the high-confidence area is weighted based on the confidence map, which further improves the reliability and decision weight of the fusion feature, and through the fusion feature driven soil erosion patch distribution extraction and intensity inversion, the spatial pattern recognition and evolution trend analysis of high precision in soil erosion dynamic monitoring are realized.

[0055] 7. By mapping the optical remote sensing features and synthetic aperture radar remote sensing features processed by multi-scale residual structure and dynamic spatial attention mechanism to a unified shared semantic feature space, and introducing semantic consistency loss function and cross-modal mutual information maximization mechanism for joint optimization, the class distinction and modal retention of shared semantic features of different modalities are effectively improved, realizing the consistent expression of optical remote sensing features and synthetic aperture radar remote sensing features in the shared space, laying a solid semantic foundation for guided learning distillation and fusion feature generation. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is a flowchart of a soil erosion monitoring method of multi-source remote sensing data fusion disclosed by the embodiments of the present application;

[0057] Figure 2 is a module schematic diagram of a soil erosion monitoring device of multi-source remote sensing data fusion disclosed by the embodiments of the present application;

[0058] Figure 3 is a structural schematic diagram of an electronic device disclosed by the embodiments of the present application.

[0059] Explanation of reference signs: 201, acquisition module; 202, processing module; 203, output module; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. DETAILED DESCRIPTION

[0060] In order for those skilled in the art to better understand the technical solutions in the specification, the technical solutions in the embodiments of the specification will be described clearly and completely in conjunction with the drawings in the embodiments of the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments.

[0061] In the description of the embodiments of the present application, the words such as "for example" or "for instance" are used to indicate an example, an illustration or an illustration. Any embodiment or design scheme described as "for example" or "for instance" in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concept in a specific manner.

[0062] In the description of the embodiments of the present application, the term "a plurality of" means two or more. For example, a plurality of systems refers to two or more systems, and a plurality of screen terminals refers to two or more screen terminals. In addition, the terms "first", "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more features. The terms "include", "contain", "have" and their variants mean "include but are not limited to", unless otherwise specifically emphasized.

[0063] Soil erosion monitoring dynamically observes the erosion, transport and deposition processes of surface soil under the action of rainfall, runoff, wind or human disturbance by systematic means, identifies the spatial range, evolution trend and intensity change of soil erosion by combining ground investigation and remote sensing data analysis, supports ecological risk assessment and management decision, and remote sensing monitoring mainly relies on optical remote sensing measurement and radar remote sensing measurement. Optical remote sensing measurement extracts vegetation, bare land and humidity characteristics by capturing visible light and infrared band reflection, but is disturbed by climate conditions. Radar remote sensing measurement can realize all-weather observation by inverting surface roughness and micro-topographic change through microwave scattering. The two are the key to improve the monitoring accuracy by synergistic integration. However, the existing multi-source fusion algorithm generally stays at the pixel level superposition and empirical normalization, lacks deep modeling of the physical response mechanism and error characteristics of optical remote sensing data and synthetic aperture radar remote sensing data, resulting in information redundancy, modal conflict and feature recognition error. It is urgent to build a cross-modal physical response consistency modeling and semantic shared feature optimization mechanism to improve the physical rationality and semantic accuracy of the fusion monitoring system.

[0064] The embodiment discloses a soil erosion monitoring method for multi-source remote sensing data fusion, referring to Figure 1 , comprising the following steps S110-S140:

[0065] S110, based on the collected optical remote sensing data and synthetic aperture radar data for monitoring landform area, a cross-modal physical response consistency mapping relationship is established, and the optical remote sensing data and synthetic aperture radar remote sensing data are subjected to consistency correction processing of physical quantity level, response scale and radiation dynamic range through a unified physical scale standardization mechanism.

[0066] The soil erosion monitoring method based on multi-source remote sensing data fusion disclosed by the embodiments of the application is applied to a server. The server includes but is not limited to electronic devices such as a mobile phone, a tablet computer, a wearable device, a PC (Personal Computer), and the like, and can also be a background server running a soil erosion monitoring method based on multi-source remote sensing data fusion. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0067] In a possible implementation, a cross-modal physical response consistency mapping relationship is established based on the collected optical remote sensing data and synthetic aperture radar data for the monitored landform region. Specifically, based on the ground object type division result of the monitored landform region, a ground object spectral reflectance characteristic model of the optical remote sensing data and a ground object electromagnetic scattering response model of the synthetic aperture radar remote sensing data are respectively constructed to clearly define the physical response mechanism of different ground object categories under different remote sensing data. Based on the ground object spectral reflectance characteristic model and the ground object electromagnetic scattering response model, a mathematical mapping relationship of the optical remote sensing data and the synthetic aperture radar remote sensing data in the physical quantity level, the response scale, and the radiation dynamic range is established, so that the observation responses of the optical remote sensing data and the synthetic aperture radar remote sensing data to the same ground object category have alignment.

[0068] Specifically, first, based on the ground object type division result of the monitored landform region, existing ground survey data, classification samples, or historical remote sensing image data are used to finely divide the ground surface in the monitored landform region by using an object-oriented classification method or a supervised learning classification method, and ground object categories including vegetation coverage, bare soil, rock exposure, water body, farmland, and construction land are identified, thereby providing a standardized ground object category basis for subsequent modeling, and ensuring that the ground object type division has spatial connectivity, consistency, and time sequence stability, so as to ensure the accuracy and representativeness of the ground object response modeling.

[0069] Subsequently, after the ground object type division is completed, for each ground object category, a ground object spectral reflectance characteristic model is constructed based on the multispectral band reflectance information of the optical remote sensing data. The model forms a parameterized description for characterizing the optical remote sensing reflection behavior of different ground object categories by statistically calculating the reflectance mean, standard deviation, and spectral shape parameters of samples of each category at different bands, and introduces auxiliary features such as vegetation index, bare land index, and humidity index to enrich the distinguishability of the ground object spectral reflectance characteristic model. A ground object electromagnetic scattering response model is constructed based on the backscattering intensity, polarization ratio, and coherence parameters of the synthetic aperture radar remote sensing data. The model takes the ground object category as a unit to record the backscattering characteristics and variation trend under VV polarization, VH polarization, or dual polarization combination, and considers the incident angle sensitivity and the ground surface roughness modulation effect to form a complete ground object electromagnetic scattering response parameter set.

[0070] Then, based on the ground object spectral reflectance characteristic model and the ground object electromagnetic scattering response model, the response curve shape, physical magnitude range and scale difference of different ground object categories in optical remote sensing data and synthetic aperture radar remote sensing data are analyzed, the optical remote sensing data reflectance characteristic space and the synthetic aperture radar remote sensing data scattering intensity characteristic space are mathematically mapped by constructing a multivariate mapping function, and the conversion relationship between the two in the physical magnitude, response scale and radiation dynamic range is established, wherein the mapping function can adopt polynomial regression, local weighted regression or deep regression network to balance the non-linear relationship modeling and high-dimensional feature matching accuracy.

[0071] Finally, according to the mathematical mapping relationship, the optical remote sensing data and the synthetic aperture radar remote sensing data in the monitored landform area are respectively subjected to physical magnitude normalization processing, response scale alignment processing and radiation dynamic range standardization processing, so that the physical response characteristics of the same ground object category in the optical remote sensing data and the synthetic aperture radar remote sensing data are aligned, ensuring the physical consistency and semantic compatibility when the subsequent shared semantic feature extraction, fusion feature generation and soil erosion monitoring model input, thereby significantly improving the overall accuracy and stability of the fusion system.

[0072] In one possible implementation, the optical remote sensing data and the synthetic aperture radar remote sensing data are subjected to consistency correction processing of physical magnitude, response scale and radiation dynamic range by a unified physical scale standardization mechanism, specifically including: according to the mathematical mapping relationship, designing a unified physical scale standardization mechanism to normalize the optical remote sensing data and the synthetic aperture radar remote sensing data to a unified physical dimension and dynamic range, eliminating the observation scale deviation caused by the difference in sensor characteristics; based on the standardized optical remote sensing data and the standardized synthetic aperture radar remote sensing data, performing physical consistency checking on each corresponding ground surface unit in the monitored landform area, ensuring that the consistency of the physical response meets the preset threshold requirement, and performing mask processing on the data area that does not meet the consistency requirement, and obtaining the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data after processing, respectively.

[0073] Specifically, according to the mathematical mapping relationship established between the optical remote sensing data and the synthetic aperture radar remote sensing data, a unified physical scale standardization mechanism is designed, and by determining a unified physical dimension reference and a radiation dynamic range reference, the reflectivity characteristics of the optical remote sensing data and the backscattering characteristics of the synthetic aperture radar remote sensing data are subjected to physical order of magnitude standardization processing, that is, the respective characteristic values are linearly transformed to a set unified physical interval, the reflectivity value of the optical remote sensing data is standardized to the interval of 0 to 1, and the backscattering coefficient of the synthetic aperture radar remote sensing data is standardized to the interval of -1 to 1, and at the same time, the scale response curve is normalized according to the mapping relationship, so that the response dynamic range of different remote sensing data to the same ground object category remains consistent, and the dimension deviation and scale distortion caused by different observation mechanisms and calibration standards of different sensors are completely eliminated.

[0074] After the standardization processing is completed, for each corresponding ground surface unit in the monitored landform area, the standardized optical remote sensing data feature vector and the standardized synthetic aperture radar remote sensing data feature vector are extracted, the similarity index of the two groups of feature vectors in the unified feature space, such as the cosine similarity, the Euclidean distance or the Mahalanobis distance, is calculated, and is compared with a preset consistency threshold value. If the similarity index meets the set threshold requirement, it is considered that the optical remote sensing data and the synthetic aperture radar remote sensing data of the ground surface unit reach the consistency standard in physical response.

[0075] For the ground surface unit that does not meet the physical consistency threshold requirement, a mask processing operation is performed, and the corresponding area is marked as invalid or high uncertainty area in the subsequent processing flow, so as to avoid interference of abnormal response data on shared semantic feature extraction and fusion feature generation, and at the same time, small range missing areas can be repaired according to the neighborhood interpolation or local reconstruction strategy, so as to improve the overall data continuity and spatial integrity.

[0076] Through the above unified physical scale standardization mechanism and physical consistency verification processing, the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data are respectively obtained, and the two types of corrected remote sensing data will be used as the standard input for subsequent multi-modal feature extraction, guided learning distillation and fusion feature generation, so as to ensure that the subsequent water and soil loss monitoring system carries out high-precision dynamic analysis and evolution trend evaluation based on physically consistent basic data.

[0077] S120, after the physical scale standardization processing is completed, shared semantic features are extracted from the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data, and a multi-scale residual structure and a dynamic spatial attention mechanism are used to compensate the structure and adjust the sensitive weight of the optical remote sensing features and the synthetic aperture radar remote sensing features.

[0078] In a possible implementation, after the physical scale standardization processing is completed, the corrected optical remote sensing data and the synthetic aperture radar remote sensing data are extracted for shared semantic features, and a multi-scale residual structure and a dynamic spatial attention mechanism are used to compensate the structure and adjust the sensitive weight of the optical remote sensing features and the synthetic aperture radar remote sensing features. Specifically, the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data are subjected to multi-scale feature extraction, respectively, and the optical remote sensing features and the synthetic aperture radar remote sensing features are extracted through convolution kernels of different scales. In the process of extracting the optical remote sensing features and the synthetic aperture radar remote sensing features, a multi-scale residual structure is superimposed, and a cross-scale residual connection method is used to retain key boundary information and structural detail information in each scale feature. Based on the optical remote sensing features and the synthetic aperture radar remote sensing features output by the multi-scale residual structure, a dynamic spatial attention mechanism is introduced, and position-sensitive weight maps are generated for the optical remote sensing feature channels and the synthetic aperture radar remote sensing feature channels, respectively, to dynamically weight and modulate the importance of features at different spatial positions. Based on the weighting results of the dynamic spatial attention mechanism, the optical remote sensing features and the synthetic aperture radar remote sensing features are enhanced in sensitive areas and suppressed in redundant areas. The optical remote sensing features and the synthetic aperture radar remote sensing features subjected to the multi-scale residual structure compensation and the dynamic spatial attention mechanism weighting adjustment are taken as shared semantic features.

[0079] Specifically, first, the optical remote sensing data and the synthetic aperture radar remote sensing data subjected to the physical scale standardization processing are input into a multi-scale feature extraction module. The multi-scale feature extraction module uses convolution kernel structures with different receptive fields to independently extract local fine-grained features and global coarse-grained features at each scale, thereby obtaining optical remote sensing features and synthetic aperture radar remote sensing features at multiple scales, so as to ensure that different spatial scale change patterns are fully captured and the spatial diversity and structural integrity of feature representation are improved.

[0080] In the multi-scale feature extraction process, a multi-scale residual structure is superimposed. The multi-scale residual structure uses a cross-scale residual connection mechanism to jumpingly connect and add residuals of low-level scale features and high-level scale features, thereby avoiding gradient disappearance and feature information loss caused by an increase in the number of convolution layers in the feature extraction process, effectively retaining key boundary detail information and surface microstructure features, and strengthening the continuity expression capability of the optical remote sensing features and the synthetic aperture radar remote sensing features in the spatial boundary area.

[0081] On the basis of the optical remote sensing features and synthetic aperture radar remote sensing features output by the multi-scale residual structure, a dynamic spatial attention mechanism is introduced. According to the local feature response strength and global feature dependency of each spatial position, the dynamic spatial attention mechanism generates a position-sensitive weight map for the optical remote sensing feature channel and the synthetic aperture radar remote sensing feature channel, respectively, dynamically adjusts the importance of different spatial regions in the feature fusion process, and adaptively highlights the key change regions and suppresses the redundant irrelevant regions.

[0082] According to the position-sensitive weight map generated by the dynamic spatial attention mechanism, the sensitive region enhancement and the redundant region suppression are performed on the optical remote sensing features and the synthetic aperture radar remote sensing features, respectively. The feature response is weighted and amplified in the sensitive region, and the feature amplitude is appropriately attenuated in the redundant region, so that the final feature representation is more focused on the water and soil loss patches, the gully erosion channels and the change sensitive areas, and the perception ability of the feature to the slight disturbance and the structural change is significantly improved.

[0083] The optical remote sensing features and the synthetic aperture radar remote sensing features that have undergone multi-scale residual structure compensation and dynamic spatial attention mechanism weighted adjustment are subjected to feature-level fusion processing to form shared semantic features with complete structure, clear boundary, consistent semantics and modal coordination. This lays a high-quality feature foundation for the subsequent guided learning distillation network optimization and fusion feature generation, and further improves the spatial resolution and classification accuracy of the soil erosion monitoring task.

[0084] In the process of extracting shared semantic features, S130, the guided learning distillation network is used to correct the feature of the confused area in the optical remote sensing data by the structural texture features of the synthetic aperture radar remote sensing data, suppress the modal conflict features and improve the boundary recognition ability.

[0085] In a possible implementation, in the process of extracting shared semantic features, the structural texture features of synthetic aperture radar remote sensing data are used to correct the features of confused areas in optical remote sensing data, suppress modal conflict features, and improve boundary recognition capability by guiding learning distillation network, and specifically includes: based on the synthetic aperture radar remote sensing data after correction processing, structural texture features are extracted to extract the ground fine-grained texture pattern, and a synthetic aperture radar structural texture feature map is generated; based on the optical remote sensing data after correction processing, an optical remote sensing confused area feature map is extracted, wherein the confused area is a spatial position with fuzzy boundary or overlapping ground object categories; a guided learning distillation network is constructed, the synthetic aperture radar structural texture feature map is used as a guide signal, and the optical remote sensing confused area feature map is used as a learning signal, a structural preservation loss function and a modal consistency loss function are jointly optimized, and the optical remote sensing confused area feature map is driven to gradually approach the spatial distribution pattern of the synthetic aperture radar structural texture feature map; in the guided learning distillation process, the boundary position of the confused area is corrected by local alignment in the distillation network, while the modal conflict features and artifact features in the optical remote sensing confused area feature map are suppressed, and the spatial boundary definition and semantic distinction are improved; the optical remote sensing features after correction by the guided learning distillation network are obtained as processed optical remote sensing features, and the synthetic aperture radar remote sensing features after correction by the guided learning distillation network are obtained as processed synthetic aperture radar remote sensing features.

[0086] Specifically, first, based on the synthetic aperture radar remote sensing data after physical scale standardization and physical consistency verification, a structural texture feature extraction module is used to extract ground fine-grained texture patterns at different spatial scales by using texture operators such as local variance, structural tensor features, Gabor filter response, or small-scale direction gradient histogram, so as to generate a synthetic aperture radar structural texture feature map that can reflect the roughness change of the ground, the erosion groove morphology, and the continuity of the ground object boundary, as a high-credibility spatial structure reference signal in the guided learning stage.

[0087] Subsequently, based on the optical remote sensing data after physical scale standardization and physical consistency verification, a boundary fuzziness detection algorithm and a category overlap analysis method are used to extract an optical remote sensing confused area feature map, wherein the confused area is defined as a position with unclear spectral features and uncertain category discrimination at the junction of vegetation coverage, bare ground exposure, water body boundary, and human activity interference area, providing an attention area for subsequent distillation guidance.

[0088] Then, a guided learning distillation network is constructed, the synthetic aperture radar structural texture feature map is taken as a guided signal, the optical remote sensing confusion area feature map is taken as a learning signal, a structure preservation loss function is introduced in the network training process to maintain the consistency of the surface space structure, a modal consistency loss function is introduced to minimize the difference between the optical remote sensing features and the synthetic aperture radar remote sensing features in the spatial distribution mode, and a joint optimization mechanism is used to drive the optical remote sensing confusion area features to gradually approach the synthetic aperture radar structural texture features, thereby enhancing the boundary clarity and the separability of the ground object categories.

[0089] After that, in the guided learning distillation process, a local alignment module is arranged in the distillation network, the local alignment module performs local similarity calculation and alignment optimization on the local neighborhood of each spatial unit in the optical remote sensing confusion area feature map, further corrects the boundary position offset through the local alignment operation, and removes the modal conflict features and artifact features caused by sensor differences or observation errors, thereby effectively improving the spatial boundary clarity and the category semantic distinguishability.

[0090] Finally, the processed optical remote sensing features are obtained through the corrected optical remote sensing features of the guided learning distillation network, and the processed synthetic aperture radar remote sensing features are obtained through the corrected synthetic aperture radar remote sensing features of the guided learning distillation network, which are used as the basis input for subsequent fusion feature generation and soil and water loss monitoring tasks, so as to ensure that the multi-source remote sensing data fusion features have unified spatial structure expression and stable physical consistency support.

[0091] In a possible implementation, after completing the physical scale normalization processing, the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data are extracted to obtain shared semantic features, and the method specifically further includes: mapping the optical remote sensing features and the synthetic aperture radar remote sensing features processed by the multi-scale residual structure and the dynamic spatial attention mechanism to the same shared semantic feature space to obtain shared optical remote sensing features and shared radar remote sensing features; defining a standardized category center vector for each ground object category in the shared semantic feature space, and constraining the shared semantic features of the same category samples to be close to the same category center vector and the shared semantic features of different category samples to be far away from other category centers by constructing a semantic consistency loss function; calculating the mutual information amount of the shared optical remote sensing features and the shared radar remote sensing features in the shared semantic feature space, and maximizing the mutual information between the shared optical remote sensing features and the shared radar remote sensing features by using a cross-modal mutual information maximization mechanism; jointly optimizing the semantic consistency loss function and the cross-modal mutual information maximization loss function in the process of calculating the mutual information amount, and maintaining the balance of the shared semantic features in the category distinguishability and the modal reservation; and respectively optimizing the shared optical remote sensing features and the shared radar remote sensing features.

[0092] Specifically, after the physical scale standardization processing and the physical consistency verification of the optical remote sensing data and the synthetic aperture radar remote sensing data, the data are respectively subjected to multi-scale residual structure compensation and dynamic spatial attention mechanism weighting processing, and then input into a shared semantic encoder module. The shared semantic encoder maps the optical remote sensing features and the synthetic aperture radar remote sensing features to the same shared semantic feature space through a set of weight-shared deep feature mapping networks, so as to ensure that the different modal features in the shared semantic feature space have unified semantic interpretation and similar spatial distribution structure, thereby forming shared optical remote sensing features and shared radar remote sensing features.

[0093] In the shared semantic feature space, a standardized class center vector is defined for each type of ground object according to the ground object type division result of the monitored landform area. Each class center vector represents the ideal position of the class in the shared semantic feature space. By constructing a semantic consistency loss function, an attractive constraint is applied to the shared optical remote sensing features and the shared radar remote sensing features of the same class, so that the feature vectors are close to the corresponding class center vectors. At the same time, an exclusion constraint is applied to the features of different classes to ensure that the shared semantic features of different classes maintain sufficient distance in space and improve the class discrimination.

[0094] The mutual information of the shared optical remote sensing features and the shared radar remote sensing features in the shared semantic feature space is calculated respectively. The mutual information reflects the size and mutual dependence of the shared information between different modal features. A cross-modal mutual information maximization mechanism is adopted to maximize the mutual information between the shared optical remote sensing features and the shared radar remote sensing features, thereby strengthening the alignment degree and complementary information utilization efficiency of the two modalities in the shared semantic feature space, and enhancing the consistency and collaborative expression ability of modal fusion.

[0095] During the training process, the semantic consistency loss function and the cross-modal mutual information maximization loss function are jointly optimized. While maintaining the aggregation of samples of the same class and the discrimination of samples of different classes, the modal consistency and complementary information retention ability between the shared optical remote sensing features and the shared radar remote sensing features are also improved. Through joint loss optimization, the shared semantic features have high class discrimination and high modal retention at the same time, ensuring the stability and reliability of feature expression in the subsequent fusion feature generation and guided learning distillation process.

[0096] The shared optical remote sensing features after the above joint optimization are used as the guide signal, and the shared radar remote sensing features after the above joint optimization are used as the learning signal, which are respectively input into the guided learning distillation network to further correct the features of the confusion area in the optical remote sensing data, suppress the modal conflict features, and improve the spatial boundary clarity of the finally generated features and the discrimination performance of the soil and water loss monitoring model.

[0097] S140, after generating the fusion features according to the processed optical remote sensing features and the processed synthetic aperture radar remote sensing features, performing confidence elimination processing and confidence weighting processing on the confidence regions in the fusion feature space, and taking the fusion features as input to drive the spatial zoning extraction and quantitative evaluation of soil erosion, so as to realize the monitoring and evaluation of the soil erosion distribution pattern and the spatio-temporal evolution trend of the geomorphic region.

[0098] In a possible implementation, after generating the fusion features according to the processed optical remote sensing features and the processed synthetic aperture radar remote sensing features, performing confidence elimination processing and confidence weighting processing on the confidence regions in the fusion feature space, and taking the fusion features as input to drive the spatial zoning extraction and quantitative evaluation of soil erosion, so as to realize the monitoring and evaluation of the soil erosion distribution pattern and the spatio-temporal evolution trend of the geomorphic region, specifically including: taking the processed optical remote sensing features and the processed synthetic aperture radar remote sensing features as input, generating fusion features through a feature concatenation and cross-modal attention fusion mechanism, the fusion features containing optical remote sensing feature information and synthetic aperture radar remote sensing feature information; performing confidence prediction on the fusion features based on a Bayesian deep neural network to generate a confidence map corresponding to each spatial position in the fusion feature space, the confidence map representing the confidence level of the fusion features at each position; according to the spatial distribution result of the confidence map, setting a confidence elimination threshold, and performing mask processing or zero processing on the low-confidence regions in the fusion feature that are lower than the elimination threshold in the confidence map; for the high-confidence regions in the confidence map that are higher than the confidence weighting threshold, performing weighted amplification processing on the feature response according to the confidence value, to obtain a feature fusion region; inputting the fusion features of the feature fusion region into a region connectivity analysis and change detection algorithm to extract the soil erosion patch distribution pattern and perform inversion on the soil erosion intensity index, to realize the identification of the soil erosion distribution pattern and the dynamic monitoring of the spatio-temporal evolution trend of the geomorphic region.

[0099] Specifically, the guided learning distillation network corrected optical remote sensing features and synthetic aperture radar remote sensing features are respectively input into a feature fusion module, the feature fusion module adopts a feature concatenation and cross-modal attention fusion mechanism, the fusion weights of the optical remote sensing features and the synthetic aperture radar remote sensing features are dynamically adjusted by splicing in the feature dimension and introducing a cross-modal attention matrix, to generate fusion features containing optical remote sensing feature information and synthetic aperture radar remote sensing feature information, so as to realize the collaborative reinforcement and complementary feature extraction of modal information, and form a complete and unified fusion feature expression.

[0100] The confidence degree of the fusion feature is predicted based on a Bayesian deep neural network. The Bayesian deep neural network generates a confidence map corresponding to each spatial position in the fusion feature space by performing multiple forward inferences on the fusion feature and analyzing the mean and variance of the prediction output. The confidence map quantifies the reliability and uncertainty of the fusion feature at different spatial positions, providing a basis for subsequent confidence-guided feature optimization.

[0101] According to the spatial distribution result of the confidence map, a confidence elimination threshold is set. The low confidence area in the confidence map that is lower than the elimination threshold is subjected to mask processing or zero processing in the fusion feature. By eliminating the low confidence area, abnormal feature responses caused by observation errors, occlusion effects or modal conflicts are effectively eliminated, and the overall robustness and noise suppression capability of the fusion feature are improved.

[0102] For the high confidence area in the confidence map that is higher than the confidence weighted threshold, the fusion feature response is subjected to weighted amplification processing according to the confidence value of each spatial position. The feature expression of the high confidence area is strengthened, so that it has higher weight and decision-making influence in the subsequent soil and water loss spatial partition extraction and soil and water loss quantitative evaluation process, forming a feature fusion area, and ensuring the physical reliability and spatial accuracy of the extraction result.

[0103] The fusion feature of the feature fusion area is taken as input. First, the soil and water loss patch distribution pattern is extracted by applying a region connectivity analysis and change detection algorithm. The soil and water loss unit is determined by analyzing the morphological features, change trend and texture structure of the connected region in the fusion feature. Then, combined with the soil and water loss intensity inversion model, the soil and water loss intensity index is inverted based on the fusion feature and external auxiliary factors (such as slope, rainfall intensity, etc.), realizing accurate identification of the soil and water loss distribution pattern in the monitored landform area and dynamic monitoring and quantitative evaluation of the spatio-temporal evolution trend, supporting watershed management and ecological restoration decision-making.

[0104] In one possible implementation, first, the fusion feature of the feature fusion area is taken as input, and a region connectivity analysis module is applied to extract each connected region in the fusion feature based on spatial adjacency relationship and pixel feature similarity. The region connectivity analysis module adopts eight-neighborhood connectivity discrimination rules and combines a connected component labeling algorithm to identify continuous and feature-consistent soil and water loss candidate patches in the fusion feature space, providing basic spatial units for subsequent change detection and soil and water loss unit determination.

[0105] Subsequently, on the basis of the obtained connected regions, a change detection module is applied to compare the multi-temporal fusion features of the monitored landform region, extract the change regions, and filter out the connected regions with significant dynamic evolution characteristics by calculating the change amplitude, change rate and change direction of the fusion features in the time sequence, so that the soil and water loss extraction process not only depends on the static spatial distribution, but also captures the dynamic evolution trend, and the identification ability of the loss sensitive area is improved.

[0106] Then, for the soil and water loss candidate connected regions determined after the change detection, morphological features of each connected region are extracted, including area, perimeter, shape index, compactness, fractal dimension and the like, and texture structure features are extracted, such as local variance, energy, entropy and homogeneity index. By comprehensively analyzing the morphological features and the texture structure features, it is determined whether each connected region meets the spatial morphological feature standard of the soil and water loss unit, and the noise spots or non-erosion change regions are removed.

[0107] After that, the connected regions determined as the soil and water loss unit are taken as the soil and water loss spatial partition result, and the soil and water loss intensity is further quantified based on the inversion model. The soil and water loss intensity inversion model takes the fusion features as the main input, combines auxiliary factors such as slope, aspect, rainfall intensity, vegetation coverage index, soil type and the like, and calculates the unit area soil loss amount index through a weighted regression model, a machine learning regression model or a process model based on the modified RUSLE equation, to form a soil and water loss intensity spatial distribution map.

[0108] Finally, the soil and water loss patch distribution pattern and the soil and water loss intensity index are superimposed to construct a soil and water loss comprehensive evolution map of the monitored landform region, the expansion trend, evolution rate and potential risk area of the loss unit are evaluated through time series analysis, the watershed management planning, ecological restoration design and soil and water conservation measure optimization decision are supported, and the spatial accuracy, dynamic responsiveness and application guidance of the monitoring system are ensured.

[0109] The embodiment also discloses a soil and water loss monitoring device based on multi-source remote sensing data fusion, referring to Figure 2 , comprising an acquisition module 201, a processing module 202 and an output module 203, and the device is used for executing any one of the soil and water loss monitoring methods based on multi-source remote sensing data fusion as described above, wherein:

[0110] The acquisition module 201 is used for establishing a cross-modal physical response consistency mapping relationship based on the collected optical remote sensing data and synthetic aperture radar data for the monitored landform region, and performing consistency correction processing on the physical magnitude, response scale and radiation dynamic range of the optical remote sensing data and the synthetic aperture radar remote sensing data through a unified physical scale standardization mechanism.

[0111] The processing module 202 is configured to extract shared semantic features from the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data after completing the physical scale standardization processing, and perform structural compensation and sensitive weight adjustment on optical remote sensing features and synthetic aperture radar remote sensing features by using a multi-scale residual structure and a dynamic spatial attention mechanism, where the optical remote sensing features are features corresponding to the corrected optical remote sensing data, and the synthetic aperture radar remote sensing features are features corresponding to the corrected synthetic aperture radar remote sensing data.

[0112] The processing module 202 is configured to correct features of confused areas in the optical remote sensing data by guiding a learning distillation network to synthesize structural texture features of the synthetic aperture radar remote sensing data during the extraction of the shared semantic features, suppress modal conflict features, and improve boundary recognition capability.

[0113] The output module 203 is configured to perform elimination processing and confidence weighting processing on confidence regions in a fusion feature space after generating the fusion features from the processed optical remote sensing features and the processed synthetic aperture radar remote sensing features, and drive water and soil loss spatial partition extraction and water and soil loss quantitative evaluation by taking the fusion features as input, so as to realize monitoring and evaluation of water and soil loss distribution patterns and spatio-temporal evolution trends of the monitored landform region.

[0114] In a possible implementation, the processing module 202 is configured to construct a ground object spectral reflectance characteristic model of the optical remote sensing data and a ground object electromagnetic scattering response model of the synthetic aperture radar remote sensing data based on a ground object type division result of the monitored landform region, so as to clarify physical response mechanisms of different ground object categories under different remote sensing data.

[0115] The processing module 202 is configured to establish a mathematical mapping relationship between the optical remote sensing data and the synthetic aperture radar remote sensing data in terms of physical quantity level, response scale, and radiation dynamic range based on the ground object spectral reflectance characteristic model and the ground object electromagnetic scattering response model, so that the optical remote sensing data and the synthetic aperture radar remote sensing data have alignment in observation responses to the same ground object category.

[0116] In a possible implementation, the processing module 202 is configured to design a unified physical scale standardization mechanism according to the mathematical mapping relationship, normalize the optical remote sensing data and the synthetic aperture radar remote sensing data to a unified physical dimension and dynamic range respectively, and eliminate observation scale deviation caused by differences in sensor characteristics.

[0117] The output module 203 is configured to perform physical consistency verification on each corresponding ground surface unit in the monitored landform area based on the normalized optical remote sensing data and the normalized synthetic aperture radar remote sensing data, to ensure that the consistency of the physical response meets a preset threshold requirement, and to perform mask processing on data regions that do not meet the consistency requirement, to obtain corrected optical remote sensing data and corrected synthetic aperture radar remote sensing data respectively after processing.

[0118] In a possible implementation, the acquisition module 201 is configured to perform multi-scale feature extraction on the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data respectively, and to extract optical remote sensing features and synthetic aperture radar remote sensing features by using convolution kernels of different scale sizes.

[0119] The processing module 202 is configured to superimpose a multi-scale residual structure during the extraction of the optical remote sensing features and the synthetic aperture radar remote sensing features, and to reserve key boundary information and structural detail information in the features of each scale by using a cross-scale residual connection mode.

[0120] The processing module 202 is configured to introduce a dynamic spatial attention mechanism on the basis of the optical remote sensing features and the synthetic aperture radar remote sensing features output by the multi-scale residual structure, to generate a position-sensitive weight map for each of the optical remote sensing feature channel and the synthetic aperture radar remote sensing feature channel, and to dynamically weight and modulate the importance of features at different spatial positions.

[0121] The processing module 202 is configured to perform sensitive region enhancement and redundant region suppression on the optical remote sensing features and the synthetic aperture radar remote sensing features on the basis of the weighting result of the dynamic spatial attention mechanism.

[0122] The processing module 202 is configured to take the optical remote sensing features and the synthetic aperture radar remote sensing features that have been compensated by the multi-scale residual structure and weighted and adjusted by the dynamic spatial attention mechanism as shared semantic features.

[0123] In a possible implementation, the acquisition module 201 is configured to extract a ground surface fine-grained texture pattern by using structural texture feature extraction based on the corrected synthetic aperture radar remote sensing data, to generate a synthetic aperture radar structural texture feature map.

[0124] The acquisition module 201 is configured to extract an optical remote sensing confusion area feature map based on the corrected optical remote sensing data, wherein the confusion area is a spatial position with fuzzy boundaries or overlapped ground object categories.

[0125] The processing module 202 is configured to construct a guided learning distillation network, synthesize a synthetic aperture radar structural texture feature map as a guided signal, and synthesize an optical remote sensing confusion area feature map as a learning signal. The guided learning distillation network is jointly optimized by using a structure preserving loss function and a modal consistency loss function, so as to drive the optical remote sensing confusion area feature map to gradually approach the spatial distribution mode of the synthetic aperture radar structural texture feature map.

[0126] The processing module 202 is configured to correct the boundary position of the confusion area and suppress the modal conflict features and artifact features in the optical remote sensing confusion area feature map in the guided learning distillation process by using local alignment inside the distillation network, so as to improve the spatial boundary definition and semantic distinguishability. The optical remote sensing feature is corrected by the guided learning distillation network to obtain a processed optical remote sensing feature. The synthetic aperture radar remote sensing feature is corrected by the guided learning distillation network to obtain a processed synthetic aperture radar remote sensing feature.

[0127] In a possible implementation, the processing module 202 is configured to input the processed optical remote sensing feature and the processed synthetic aperture radar remote sensing feature, generate a fusion feature by using a feature concatenation and a cross-modal attention fusion mechanism, and the fusion feature contains optical remote sensing feature information and synthetic aperture radar remote sensing feature information.

[0128] The processing module 202 is configured to perform confidence prediction on the fusion feature based on a Bayesian deep neural network, generate a confidence map corresponding to each spatial position in the fusion feature space, and the confidence map represents the confidence level of the fusion feature at each position.

[0129] The processing module 202 is configured to set a confidence elimination threshold according to the spatial distribution result of the confidence map, and perform mask processing or zero processing on a low-confidence area in the fusion feature, which is lower than the elimination threshold in the confidence map.

[0130] The processing module 202 is configured to perform weighted amplification processing on a high-confidence area in the confidence map, which is higher than a confidence weighting threshold, according to the confidence value of the high-confidence area, to obtain a feature fusion area.

[0131] The processing module 202 is configured to input the fusion feature of the feature fusion area into a region connectivity analysis and change detection algorithm to extract a water and soil erosion patch distribution pattern, and perform inversion on a water and soil erosion intensity index, so as to realize identification of the water and soil erosion distribution pattern and dynamic monitoring of the spatio-temporal evolution trend of the topography area.

[0132] In a possible implementation, the processing module 202 is configured to map the optical remote sensing feature and the synthetic aperture radar remote sensing feature processed by using a multi-scale residual structure and a dynamic spatial attention mechanism to a same shared semantic feature space, to obtain a shared optical remote sensing feature and a shared radar remote sensing feature.

[0133] The processing module 202 is configured to define a standardized class center vector for each ground object class in the shared semantic feature space, and constrain shared semantic features of samples of the same class to be close to the same class center vector and shared semantic features of samples of different classes to be far away from other class centers by constructing a semantic consistency loss function.

[0134] The processing module 202 is configured to calculate mutual information of the shared optical remote sensing features and the shared radar remote sensing features in the shared semantic feature space, and maximize the mutual information between the shared optical remote sensing features and the shared radar remote sensing features by using a cross-modal mutual information maximization mechanism.

[0135] The processing module 202 is configured to jointly optimize the semantic consistency loss function and the cross-modal mutual information maximization loss function in the process of calculating the mutual information, and balance the class distinguishability and the modal reservation of the shared semantic features, and optimize the shared optical remote sensing features and the shared radar remote sensing features respectively.

[0136] The processing module 202 is configured to use the optimized shared optical remote sensing features as a guide signal and the optimized shared radar remote sensing features as a learning signal.

[0137] It should be noted that the apparatus provided in the above embodiments is only used as an example to illustrate the division of the above functional modules in realizing its functions, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the above described functions. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0138] The embodiment further discloses an electronic device, which refers to Figure 3 The electronic device can include at least one processor 301, at least one communication bus 302, a user interface 303, a network interface 304, and at least one memory 305.

[0139] The communication bus 302 is configured to realize the connection and communication between the components.

[0140] The user interface 303 can include a display screen (Display) and a camera (Camera), and the optional user interface 303 can further include a standard wired interface and a wireless interface.

[0141] The network interface 304 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0142] The processor 301 can include one or more processing cores. The processor 301 connects various parts within the server through various interfaces and lines, performs various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 305, and calling data stored in the memory 305. Alternatively, the processor 301 can be implemented in at least one of a hardware form of a digital signal processing (DSP), a field-programmable gate array (FPGA), and a programmable logic array (PLA). The processor 301 can integrate a combination of one or more of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. Among them, the CPU mainly processes operating systems, user interfaces, and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; and the modem is used for processing wireless communication. It can be understood that the above-mentioned modem can also not be integrated into the processor 301, but can be realized by a separate chip.

[0143] The memory 305 can include a random access memory (RAM) and a read-only memory (ROM). Optionally, the memory includes a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 can include a program storage area and a data storage area, wherein the program storage area can store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playing function, an image playing function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area can store data involved in the above-mentioned various method embodiments, etc. The memory 305 can also be at least one storage device located away from the aforementioned processor 301. As a computer storage medium, the memory 305 can include an operating system, a network communication module, a user interface 303 module, and an application program of a kind of multi-source remote sensing data fusion soil and water loss monitoring method.

[0144] In Figure 3In the electronic device shown, the user interface 303 is mainly used to provide an interface for user input, and obtain data input by the user; and the processor 301 can be used to invoke an application program of a method for monitoring soil erosion by fusing multi-source remote sensing data stored in the memory 305, and when executed by one or more processors 301, the electronic device performs the method of one or more of the above embodiments.

[0145] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all described as a series of action combinations, but those skilled in the art should know that the present application is not limited to the order of the actions described, because according to the present application, certain steps can be performed in other order or at the same time. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0146] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0147] In several embodiments provided by the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic. The division of the units is only a logical function division. There can be another division manner for actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different parts can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical or other forms.

[0148] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0149] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically independently, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0150] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable memory. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a memory 305 and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned memory 305 includes: a U disk, a mobile hard disk, a magnetic disk or an optical disk and various program code storage media.

[0151] The present application also discloses a computer readable storage medium, which stores instructions. When executed by one or more processors 301, the instructions cause an electronic device to perform one or more methods as described in the above embodiments.

[0152] The above are only exemplary embodiments of the present disclosure, and cannot limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon considering the specification and practicing the true principles of the present disclosure. The present application is intended to cover any variations, uses or adaptive changes of the present disclosure that follow the general principles of the present disclosure and include common knowledge or conventional technical means in the art that are not described in the present disclosure. The specification and examples are only considered as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A soil erosion monitoring method based on multi-source remote sensing data fusion, characterized in that, The method comprises: Based on the collected optical remote sensing data and synthetic aperture radar remote sensing data for monitoring the geomorphic region, a cross-modal physical response consistency mapping relationship is established, and a unified physical scale standardization mechanism is used to correct the consistency of the physical quantity level, response scale and radiation dynamic range of the optical remote sensing data and the synthetic aperture radar remote sensing data; After completing the physical scale standardization processing, the shared semantic features are extracted from the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data, and a multi-scale residual structure and a dynamic spatial attention mechanism are used to compensate the structure and adjust the sensitive weight of the optical remote sensing features and the synthetic aperture radar remote sensing features, wherein the optical remote sensing features are the features corresponding to the corrected optical remote sensing data, and the synthetic aperture radar remote sensing features are the features corresponding to the corrected synthetic aperture radar remote sensing data; In the process of extracting shared semantic features, the structure texture features of the synthetic aperture radar remote sensing data are used to guide the learning of the distillation network to correct the features of the confusion area in the optical remote sensing data, suppress the modal conflict features and improve the boundary recognition ability; After generating the fusion features based on the processed optical remote sensing features and the processed synthetic aperture radar remote sensing features, the confidence regions in the fusion feature space are removed and weighted, and the fusion features are used as input to drive the water and soil loss spatial partition extraction and the water and soil loss quantitative evaluation, so as to realize the monitoring and evaluation of the water and soil loss distribution pattern and the spatio-temporal evolution trend of the monitoring geomorphic region. 2.The soil erosion monitoring method of multi-source remote sensing data fusion according to claim 1, characterized in that, The method comprises: Based on the ground object type division result of the monitoring geomorphic region, the ground object spectral reflection characteristic model of the optical remote sensing data and the ground object electromagnetic scattering response model of the synthetic aperture radar remote sensing data are constructed respectively to clarify the physical response mechanism of different ground object categories under different remote sensing data; Based on the ground object spectral reflection characteristic model and the ground object electromagnetic scattering response model, a mathematical mapping relationship between the optical remote sensing data and the synthetic aperture radar remote sensing data in the physical quantity level, response scale and radiation dynamic range is established, so that the observation responses of the optical remote sensing data and the synthetic aperture radar remote sensing data to the same ground object category have alignment. 3.The soil erosion monitoring method of multi-source remote sensing data fusion according to claim 2, characterized in that, The method comprises: According to the mathematical mapping relationship, a unified physical scale standardization mechanism is designed to normalize the optical remote sensing data and the synthetic aperture radar remote sensing data to a unified physical dimension and dynamic range, eliminating the observation scale deviation caused by the difference in sensor characteristics. The physical consistency of each corresponding ground surface unit in the monitored landform area is verified based on the normalized optical remote sensing data and the normalized synthetic aperture radar remote sensing data, to ensure that the consistency of the physical response meets the preset threshold requirement, and the data area that does not meet the consistency requirement is subjected to mask processing, and the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data are obtained after the processing. 4.The soil erosion monitoring method of multi-source remote sensing data fusion according to claim 1, characterized in that, After the physical scale normalization processing is completed, shared semantic features are extracted from the corrected optical remote sensing data and the synthetic aperture radar remote sensing data, and a multi-scale residual structure and a dynamic spatial attention mechanism are used to compensate the structure and adjust the sensitive weight of the optical remote sensing features and the synthetic aperture radar remote sensing features, specifically including: The corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data are subjected to multi-scale feature extraction, and the optical remote sensing features and the synthetic aperture radar remote sensing features are extracted through convolution kernels of different scale sizes; During the extraction of the optical remote sensing features and the synthetic aperture radar remote sensing features, a multi-scale residual structure is superimposed, and a cross-scale residual connection method is used to retain key boundary information and structural detail information in each scale feature; Based on the optical remote sensing features and the synthetic aperture radar remote sensing features output by the multi-scale residual structure, a dynamic spatial attention mechanism is introduced, and position-sensitive weight maps are generated for the optical remote sensing feature channels and the synthetic aperture radar remote sensing feature channels respectively, to dynamically weight and modulate the importance of features at different spatial positions; Based on the weighting results of the dynamic spatial attention mechanism, the optical remote sensing features and the synthetic aperture radar remote sensing features are subjected to sensitive region enhancement and redundant region suppression; The optical remote sensing features and the synthetic aperture radar remote sensing features subjected to multi-scale residual structure compensation and dynamic spatial attention mechanism weighting adjustment are used as the shared semantic features. 5.The soil erosion monitoring method of multi-source remote sensing data fusion according to claim 1, characterized in that, During the extraction of the shared semantic features, a guided learning distillation network is used to correct the feature of the confusion area in the optical remote sensing data based on the structural texture features of the synthetic aperture radar remote sensing data, to suppress the modal conflict features and improve the boundary recognition ability, specifically including: Based on the corrected synthetic aperture radar remote sensing data, structural texture feature extraction is used to extract the ground fine-grained texture pattern, to generate a synthetic aperture radar structural texture feature map; Based on the corrected optical remote sensing data, an optical remote sensing confusion area feature map is extracted, wherein the confusion area is a spatial position with fuzzy boundary or overlapped ground object category; The learning distillation network is constructed, the synthetic aperture radar structural texture feature map is used as a guide signal, and the optical remote sensing confusion area feature map is used as a learning signal, a structure preserving loss function and a modal consistency loss function are used for joint optimization, and the optical remote sensing confusion area feature map is driven to gradually approach the spatial distribution pattern of the synthetic aperture radar structural texture feature map. In the guided learning distillation process, the boundary position of the confusion area is corrected through local alignment within the distillation network, while the modal conflict features and artifact features in the optical remote sensing confusion area feature map are suppressed, and the spatial boundary clarity and semantic distinguishability are improved; The optical remote sensing feature is corrected by the guided learning distillation network to obtain the processed optical remote sensing feature, and the synthetic aperture radar remote sensing feature is corrected by the guided learning distillation network to obtain the processed synthetic aperture radar remote sensing feature. 6.The soil erosion monitoring method of multi-source remote sensing data fusion according to claim 1, characterized in that, After generating the fusion feature based on the processed optical remote sensing feature and the processed synthetic aperture radar remote sensing feature, the confidence region in the fusion feature space is removed and weighted, and the fusion feature is input to drive the soil and water loss spatial partition extraction and soil and water loss quantitative evaluation, thereby realizing the monitoring and evaluation of the soil and water loss distribution pattern and the spatio-temporal evolution trend of the monitored landform area, specifically including: The processed optical remote sensing feature and the processed synthetic aperture radar remote sensing feature are input to generate the fusion feature through feature concatenation and cross-modal attention fusion mechanism, and the fusion feature contains optical remote sensing feature information and synthetic aperture radar remote sensing feature information; Based on the Bayesian deep neural network, the confidence of the fusion feature is predicted to generate a confidence map corresponding to each spatial position in the fusion feature space, and the confidence map represents the confidence level of the fusion feature at each position; According to the spatial distribution result of the confidence map, a confidence removal threshold is set, and a low confidence region below the removal threshold in the confidence map is masked or set to zero in the fusion feature; For a high confidence region in the confidence map above a confidence weighting threshold, the feature response is amplified by weighting according to the confidence value, to obtain a feature fusion region; The fusion feature of the feature fusion region is input to extract the soil and water loss patch distribution pattern based on regional connectivity analysis and change detection algorithm, and to inverse the soil and water loss intensity index, to realize the identification of the soil and water loss distribution pattern and the dynamic monitoring of the spatio-temporal evolution trend of the monitored landform area. 7.The soil erosion monitoring method of multi-source remote sensing data fusion according to claim 5, characterized in that, After completing the physical scale normalization processing, the corrected optical remote sensing data and synthetic aperture radar remote sensing data are extracted to obtain shared semantic features, specifically including: The optical remote sensing feature and the synthetic aperture radar remote sensing feature processed by the multi-scale residual structure and the dynamic spatial attention mechanism are mapped to the same shared semantic feature space to obtain shared optical remote sensing features and shared radar remote sensing features; In the shared semantic feature space, a standardized class center vector is defined for each ground object class, and a semantic consistency loss function is constructed to constrain the shared semantic features of samples of the same class to be close to the same class center vector, and the shared semantic features of samples of different classes to be far away from other classes; The mutual information of the shared optical remote sensing feature and the shared radar remote sensing feature in the shared semantic feature space is calculated, and a cross-modal mutual information maximization mechanism is used to maximize the mutual information between the shared optical remote sensing feature and the shared radar remote sensing feature; The semantic consistency loss function and the cross-modal mutual information maximization loss function are jointly optimized in the calculation of the mutual information amount, while keeping the balance of the shared semantic features in the aspects of class discrimination and modality reservation, and the shared optical remote sensing features and the shared radar remote sensing features are optimized respectively. The optimized shared optical remote sensing features are taken as the guide signal, and the optimized shared radar remote sensing features are taken as the learning signal.

8. A soil erosion monitoring device of multi-source remote sensing data fusion, characterized in that, The device is used for executing the method for monitoring soil and water loss by fusing multi-source remote sensing data, and the device comprises an acquisition module (201), a processing module (202), and an output module (203). The acquisition module (201) is configured to establish a cross-modal physical response consistency mapping relationship based on the collected optical remote sensing data and synthetic aperture radar remote sensing data for monitoring a landform region, and perform consistency correction processing on the optical remote sensing data and the synthetic aperture radar remote sensing data in terms of physical quantity level, response scale, and radiation dynamic range through a unified physical scale standardization mechanism. The processing module (202) is configured to extract shared semantic features from the corrected optical remote sensing data and the corrected synthetic aperture radar remote sensing data after completing the physical scale standardization processing, and perform structure compensation and sensitive weight adjustment on optical remote sensing features and synthetic aperture radar remote sensing features by using a multi-scale residual structure and a dynamic spatial attention mechanism, wherein the optical remote sensing features are features corresponding to the corrected optical remote sensing data, and the synthetic aperture radar remote sensing features are features corresponding to the corrected synthetic aperture radar remote sensing data. The processing module (202) is configured to correct features of a confused area in the optical remote sensing data by using structure texture features of the synthetic aperture radar remote sensing data to suppress modality conflict features and improve boundary recognition capability during the extraction of the shared semantic features by using a guided learning distillation network. The output module (203) is configured to perform confidence region elimination processing and confidence weighting processing on a fusion feature space after generating the fusion feature according to the processed optical remote sensing features and the processed synthetic aperture radar remote sensing features, and drive soil and water loss spatial zoning extraction and soil and water loss quantitative evaluation by taking the fusion feature as input, so as to realize monitoring and evaluation of a soil and water loss distribution pattern and a spatio-temporal evolution trend of the monitoring landform region.

9. An electronic device, comprising: The electronic device comprises a processor (301), a communication bus (302), a user interface (303), a network interface (304), and a memory (305), the memory (305) is configured to store instructions, the user interface (303) and the network interface (304) are configured to communicate with other devices, the communication bus (302) is configured to realize connection and communication between components in the electronic device, and the processor (301) is configured to execute the instructions stored in the memory (305) to enable the electronic device to execute the method.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores instructions that, when executed, perform the method of any of claims 1-7.

Images