Remote sensing image land cover monitoring system based on data-model coupling driving

Abstract

The application relates to the technical field of image recognition and discloses a remote sensing image land cover monitoring system based on data-model coupling driving, which comprises a multi-temporal data loading module, a change area extraction module, a time sequence consistency optimization module, a reliable change discrimination module and a structured monitoring product generation module; after multi-temporal land cover classification results are generated, the system uses time sequence consistency constraints to iteratively optimize the classification probability, suppresses pseudo-change noise generated due to single-temporal classification uncertainty, obtains time sequence consistent change results, automatically classifies by calculating the average confidence of the change area and setting a threshold, divides the results into high-confidence change and change to be checked, the process directly outputs monitoring products which are spatiotemporal consistent and are attached with confidence labels, significantly reduces invalid information which needs to be identified by humans, facilitates business departments to focus on high-confidence discovery and key checking, and thus the reliability and practical efficiency of automatic time sequence monitoring results are improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a remote sensing image land cover monitoring system based on data-model coupling. Background Technology

[0002] Image recognition technology encompasses the branch of technologies that automatically detect, identify, classify, and understand specific targets, scenes, or content in digital images or videos. The core of this technology is enabling computer systems to simulate human visual cognitive functions, analyzing and interpreting input visual data. Its overall technology systematically covers a series of processes from image feature extraction and pattern matching to advanced semantic understanding, aiming to achieve automated and intelligent identification and classification of the information contained in images.

[0003] Among them, the data-model coupled remote sensing image land cover monitoring system refers to a system specifically designed for analyzing and processing Earth observation remote sensing images. This system addresses the technical challenge of automated classification and change monitoring of land cover types by coupling high-resolution remote sensing image data with a pre-trained recognition model. This model performs pixel-level or object-level analysis of the image data, thereby determining and mapping the land cover category of each unit in the image. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a remote sensing image land cover monitoring system based on data-model coupling, which has the functions of automatically suppressing false changes and outputting graded reliability products, thus solving the problems mentioned in the background technology.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A remote sensing image land cover monitoring system based on data-model coupling includes the following processing modules connected in sequence: The multi-temporal data loading module calls the neural network parameters, and the calculation unit performs hierarchical transformation on the pixel spectral values, outputs the probability of its classification into each category, and generates a multi-temporal pixel-level category probability matrix. The change region extraction module, based on a multi-temporal pixel-level category probability matrix, compares the maximum probability category encoding of pixels in adjacent temporal phases, records pixels with different encodings, and aggregates spatially adjacent pixels with the same change category to generate a set of land cover change polygons. The temporal consistency optimization module calls the multi-temporal pixel-level category probability matrix and the set of land cover change polygons to extract the temporal category and probability of pixels within the polygons. It determines whether the probability of a single point change is lower than a set threshold or whether the number of overall state changes exceeds a set threshold. The coordinates of suspicious pixels that meet the judgment conditions and their target temporal information are added to the queue to be processed. The original probability vectors of the pixels in the queue are weighted, adjusted and normalized according to the temporal categories of the pixels before and after. The pixel categories are calculated and updated in a loop until the queue is empty, resulting in a temporally consistent land cover patch coding sequence and an optimized set of land cover change polygons. The credible change identification module calculates the average probability value of pixels within each polygon during the change phase for the optimized set of land cover change polygons, compares this average value with a preset high confidence threshold, and classifies the polygons into a high confidence change set and a change set to be verified based on the results. The structured monitoring product generation module calls the time-consistent land cover patch coding sequence, high-confidence change set, and change set to be verified, and encapsulates the data according to the format for the spatial, category, and confidence identification attributes of the patches and polygons, generating a land cover monitoring product set with confidence level labels.

[0006] Preferably, the multi-temporal pixel-level category probability matrix is ​​specifically a type of structured data, including the spatial coordinates of each pixel, the corresponding image acquisition temporal identifier, and a probability value vector representing that the pixel belongs to each predefined land cover category; The land cover change polygon set is specifically a set of geographic features, where each feature includes a geometric outline defined by a set of continuous vertex coordinates, the dominant land cover category code before the change represented by the polygon, and the dominant land cover category code after the change. The temporally consistent land cover patch coding sequence is specifically a set of temporal geographic objects, including the unique identifier of each patch, its spatial outline in different temporal phases, and the optimized dominant land cover category coding sequence arranged in chronological order. The optimized land cover change polygon set is specifically an optimized subset of the land cover change polygon set, wherein each optimized change polygon includes an optimized geometric profile, an optimized pre-change category code, an optimized post-change category code, and a state identifier used to identify that it has undergone temporal consistency optimization. Both the high-confidence change set and the change set to be verified refer to subsets divided from the optimized land cover change polygon set according to specific rules. Each subset includes a confidence level identifier and a change event attribute table corresponding to all change polygons in the subset. The attribute table includes change area, change phase, and average change confidence. The land cover monitoring product set with confidence level identifiers is specifically the final system output data package, which includes data corresponding to the time-consistent land cover patch coding sequence, vector layers of high-confidence change patches, vector layers of change patches to be verified, and product metadata files describing the data source, processing time, and coordinate system.

[0007] Preferably, the change region extraction module includes a change pixel recording submodule and a change region aggregation submodule; The variable pixel recording submodule, based on the multi-temporal pixel-level category probability matrix, reads the probability value array of pixels in adjacent temporal phases and finds the category code corresponding to the highest probability, compares whether the category codes of two temporal phases are equal, and records the spatial coordinates and the category codes of the preceding and following temporal phases for pixels with unequal category codes, and generates a variable pixel recording table. The change region aggregation submodule calls the change pixel record table, searches for other records with the same preceding and following category codes in the eight neighboring regions based on the record coordinates, and merges the record pixels that meet the conditions into continuous polygons in space to generate a set of land cover change polygons.

[0008] Preferably, the temporal consistency optimization module includes a suspicious pixel screening submodule, a pixel category iterative optimization submodule, and an optimization result integration submodule; The suspicious pixel screening submodule calls the multi-temporal pixel-level category probability matrix and the land cover change polygon set. For pixels within the polygon, it determines whether the probability of change at a single point is lower than the lower limit of classification confidence or whether the number of changes in the overall state of the polygon exceeds the state fluctuation threshold. The pixel coordinates and target temporal information that meet the conditions are added to the queue to be processed. The pixel category iterative optimization submodule, based on the queue to be processed and the multi-temporal pixel-level category probability matrix, retrieves pixel information from the queue, performs weighted adjustment and normalization on the original probability vector according to its preceding and following temporal categories, calculates its optimized category, updates it if the category changes, and adds the affected pixel information to the tail of the queue, loops until the queue is empty, and obtains the optimized category data of all pixels. The optimization result integration submodule calls the optimized category data of all pixels, aggregates spatially adjacent pixels with the same category into patches, associates each temporal category code with the spatial contour, generates a temporally consistent land cover patch code sequence, and re-detects and aggregates change areas based on this to generate an optimized land cover change polygon set.

[0009] Preferably, the credible change identification module includes a change confidence calculation submodule and a confidence grading submodule; The change confidence calculation submodule, for the optimized set of land cover change polygons, traverses each polygon in the set, extracts all pixel coordinates covered by it, reads the corresponding dominant class probability value from the multi-temporal pixel-level class probability matrix based on these coordinates and the time phase of the change, calculates the arithmetic mean of these read probability values, and obtains the average change confidence of the polygon. The confidence level submodule calls the average change confidence of polygons, compares it with the preset high confidence threshold value, classifies polygons with an average change confidence greater than the threshold into a new set, and classifies polygons with an average change confidence less than or equal to the threshold into another new set, generating a high confidence change set and a change set to be verified.

[0010] Preferably, in the suspicious pixel identification submodule, determining whether the probability value of a pixel in adjacent time phases is less than the lower limit of classification confidence is specifically performed as follows: for a marked pixel, retrieve the two specific time phases T_a and T_b involved in the change of its category, read the dominant category probability value P_a of the pixel in time phase T_a and the dominant category probability value P_b of the pixel in time phase T_b respectively, and determine whether P_a is less than the lower limit of classification confidence or whether P_b is less than the lower limit of classification confidence.

[0011] Preferably, in the iterative optimization processing submodule, when constructing the adjustment weight vector, the fixed increment of the element value with index equal to L_prev and L_next is a constant greater than zero.

[0012] Preferably, the structured monitoring product generation module includes an attribute association encapsulation submodule and a product packaging output submodule; The attribute association encapsulation submodule calls the time-consistent land cover patch coding sequence, high-confidence change set, and change set to be verified to generate a set of geographic features with attributes for the spatial geometry, category coding, and confidence identification attributes associated with patches and polygons. The product packaging and output submodule, based on a set of geographic features with attributes, encodes the geographic features according to the specified spatial data format and creates a metadata file containing data source, temporal phase and coordinate system information. After packaging, it generates a land cover monitoring product set with a reliability rating identifier.

[0013] Preferably, the multi-temporal data loading module includes a data coupling loading submodule and a pixel-level probability generation submodule; The data coupling loading submodule acquires multi-temporal high-resolution remote sensing image data of the target area, calls the stored deep neural network model weight parameter file, and loads the image data and model weight parameters into the memory of the graphics processing unit simultaneously, generating a set of loaded image data and model parameters. The pixel-level probability generation submodule, based on the loaded image data and model parameter set, performs continuous convolution and nonlinear activation operations on the multi-band spectral values ​​corresponding to each pixel in the image. Finally, it outputs a numerical array representing the probability of the pixel belonging to each predefined land cover category. After completing the calculation for all pixels, it generates a multi-temporal pixel-level category probability matrix.

[0014] Compared with existing technologies, this invention provides a remote sensing image land cover monitoring system based on data-model coupling, which has the following advantages: This remote sensing imagery land cover monitoring system, driven by data-model coupling, introduces temporal consistency constraints to iteratively optimize classification probabilities after generating multi-temporal land cover classification results. This effectively suppresses spurious change noise caused by uncertainty in single-temporal classification, producing self-consistent change results in the time dimension. Simultaneously, by calculating the average confidence level of the change area and setting a threshold for automatic classification, the system distinguishes between high-confidence changes and changes requiring verification. This process directly outputs monitoring products with strong spatiotemporal consistency and confidence labels, significantly reducing invalid information in the results that requires manual identification. This allows business departments to focus on high-confidence discovery and key verification, greatly improving the reliability and practical efficiency of automated time-series monitoring results. Attached Figure Description

[0015] Figure 1 This is a block diagram of the remote sensing image land cover monitoring system based on data-model coupling driven by the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0017] Please see Figure 1 This invention provides a remote sensing image land cover monitoring system based on data-model coupling, comprising the following processing modules connected in sequence: The multi-temporal data loading module calls the neural network parameters, and the calculation unit performs hierarchical transformation on the pixel spectral values, outputs the probability of its classification into each category, and generates a multi-temporal pixel-level category probability matrix. The change region extraction module, based on a multi-temporal pixel-level category probability matrix, compares the maximum probability category encoding of pixels in adjacent temporal phases, records pixels with different encodings, and aggregates spatially adjacent pixels with the same change category to generate a set of land cover change polygons. The temporal consistency optimization module calls the multi-temporal pixel-level category probability matrix and the set of land cover change polygons to extract the temporal category and probability of pixels within the polygons. It determines whether the probability of a single point change is lower than a set threshold or whether the number of overall state changes exceeds a set threshold. The coordinates of suspicious pixels that meet the judgment conditions and their target temporal information are added to the queue to be processed. The original probability vectors of the pixels in the queue are weighted, adjusted and normalized according to the temporal categories of the pixels before and after. The pixel categories are calculated and updated in a loop until the queue is empty, resulting in a temporally consistent land cover patch coding sequence and an optimized set of land cover change polygons. The credible change identification module calculates the average probability value of pixels within each polygon during the change phase for the optimized set of land cover change polygons, compares this average value with a preset high confidence threshold, and classifies the polygons into a high confidence change set and a change set to be verified based on the results. The structured monitoring product generation module calls the time-consistent land cover patch coding sequence, high-confidence change set, and change set to be verified, and encapsulates the data according to the format for the spatial, category, and confidence identification attributes of the patches and polygons, generating a land cover monitoring product set with confidence level labels.

[0018] This embodiment provides the overall architecture and core process of an intelligent monitoring system. The system takes multi-temporal high-resolution remote sensing images as input and uses a trained deep neural network (e.g., a convolutional neural network) to abstract and calculate the spectral features of each pixel layer by layer. It outputs the probability that each pixel belongs to a predefined land cover category such as forest, cultivated land, or construction land, forming a structured multi-temporal pixel-level category probability matrix. This matrix serves as the data foundation for all subsequent processing. Next, the system identifies pixels that have undergone category changes by comparing the most likely category of each pixel in adjacent temporal phases (e.g., 2023 and 2024). These spatially adjacent pixels with the same change type (e.g., from "cultivated land" to "construction land") are aggregated into a continuous vector polygon, forming a land cover change polygon set, thus initially outlining the change area.

[0019] To further improve the reliability of the results over time, the system performs in-depth optimization on the initially detected change areas. It examines the class stability of pixels within each change polygon over a longer time series, identifying suspicious pixels due to low single-phase classification confidence (probability value below a threshold, such as 0.6) or frequent class jumps over time (e.g., more than two changes in five consecutive phases), and adds them to a processing queue. The system uses an iterative algorithm to process this queue: for each pixel in the queue, based on its current optimal class in the phases before and after the change (i.e., phases T-1 and T+1), the original classification probability vector of that pixel in the target phase T is weighted and adjusted (e.g., increasing the probability weight consistent with the class in the phases before and after the change), and then its optimal class is recalculated. If the class is corrected, the information of neighboring pixels affected by this correction is added to the queue for chain optimization. This iterative process continues until the queue is empty, ultimately outputting a self-consistent and temporally consistent land cover patch coding sequence and an optimized set of land cover change polygons.

[0020] Next, the system performs quality assessment and classification on the optimized change results. It calculates the average classification probability of all pixels within each changed polygon at the time of the change, using this as the average change confidence level for that area. This confidence level is compared to a preset high confidence threshold (e.g., 0.75), automatically dividing the results into a high-confidence change set (directly acceptable) and a change set requiring manual verification. Finally, the system encapsulates the time-consistent land cover classification patches, the classified change vectors, and the corresponding spatial geometry, category attributes, and confidence labels according to geographic information data standards (such as GeoTIFF and GeoJSON), and adds metadata processing to generate a final land cover monitoring product set with accompanying confidence level identifiers, ready for direct business release. This embodiment, through the above-described process of coupling data processing and model optimization judgment, achieves automated output from raw imagery to high-quality, directly applicable intelligent monitoring products. Example

[0021] Based on Embodiment 1, the change region extraction module includes a change pixel recording submodule and a change region aggregation submodule.

[0022] This embodiment refines the change region extraction module. The change pixel recording submodule specifically performs the following operations: Iterates through the multi-temporal pixel-level category probability matrix. For each pair of adjacent temporal phases (e.g., T_i = May 2023, T_{i+1} = May 2024), for each pixel location, it reads its category probability vector in these two temporal phases, for example, vectors [0.1, 0.8, 0.05, 0.03, 0.02] and [0.7, 0.2, 0.05, 0.03, 0.02] (corresponding to forest land, cultivated land, construction land, water body, and bare land, respectively). The submodule finds the maximum probability value (0.8 and 0.7) in each vector and its corresponding category code (assuming cultivated land is coded as 2 and forest land as coded as 1). Compare the category codes of the two time phases (2 and 1). Since they are not equal, it is determined that a land cover change has occurred at this location. Therefore, record the planar coordinates (X, Y) of the pixel, the category code A=2 for the time phase T_i before the change, and the category code B=1 for the time phase T_{i+1} after the change. After performing this operation on all pixels, a record table of changed pixels is generated, which records the location and type of change of all changed pixels.

[0023] The change region aggregation submodule is responsible for aggregating discrete changed pixels into meaningful patches. It reads the changed pixel record table, using the first record as the seed, and searches for other records within its eight neighborhoods (east, south, west, north, southeast, northeast, southwest, and northwest). It determines whether the searched records have the exact same preceding-following temporal category code as the seed record (A=2 and B=1). If the condition is met, the two pixels are considered to belong to the same change region and spatially merged. Using this newly merged pixel as the new starting point, it continues searching and merging records satisfying the same condition in its eight neighborhoods. This process is recursively repeated until no new matching neighborhood pixels are found. Ultimately, all merged pixels form a closed polygon composed of continuous boundary points, representing an independent change patch from "cultivated land" to "woodland." After traversing all unaggregated records in the record table, a set of land cover change polygons consisting of multiple such polygons is generated. This embodiment uses a two-step method of "recording first and then aggregating" to effectively organize pixel-level subtle changes into object-level change regions with clear geographical boundaries, providing a structured foundation for subsequent analysis. Example

[0024] Based on Example 1, the temporal consistency optimization module includes a suspicious pixel screening submodule, a pixel category iterative optimization submodule, and an optimization result integration submodule.

[0025] This embodiment elaborates on the core mechanism of temporal consistency optimization. The suspicious pixel screening submodule is responsible for locating potentially unreliable changed pixels. It utilizes both a multi-temporal pixel-level category probability matrix and a set of land cover change polygons. For an identified changed polygon, the submodule extracts the category code (taking the one with the highest probability) and the corresponding maximum probability value for each phase of all pixels within it over a longer observation period (e.g., 5 consecutive years). The screening is based on two rules: First, for a single pixel, the specific two adjacent phases in which its category changed are examined. If, in either of these two phases, the probability value of the pixel belonging to its "dominant category" (i.e., the category before or after the change) is lower than a preset classification confidence threshold (e.g., 0.6), then this pixel is marked as a "low-confidence change point". Second, for the entire polygon, the number of times the dominant category of all pixels within it changes over multiple consecutive phases is counted. If the average number of changes exceeds a preset state fluctuation threshold (e.g., 2 times), then the overall state of the polygon is considered unstable, and all pixels within it are marked as "suspected noise". The coordinates of all marked pixels and their corresponding target optimization phase information are added to a processing queue.

[0026] The pixel category iterative optimization submodule is the engine for performing optimization. It iteratively retrieves a task item (containing pixel coordinates P and target time phase T) from the head of the queue. First, it obtains the original category probability vector V_orig for the pixel at time phase T from the multi-time phase pixel-level category probability matrix, for example, [0.3, 0.55, 0.1, 0.03, 0.02] (corresponding to 5 categories). At the same time, it obtains the current optimal category codes L_prev and L_next for the pixel at times T-1 and T+1 (in the early stages of iteration, this code is the original classification result; it may have been updated during iteration). Next, it constructs a weight vector W with the same number of categories, initially all values ​​are 1. The values ​​of the elements in W with indices equal to L_prev and L_next are increased by a fixed increment Δ (e.g., Δ=0.3). If L_prev and L_next are both category 2 (farmland), then W becomes [1, 1, 1.3, 1, 1]. Then, the adjusted probability vector V_adj = W ⊙ V_orig (⊙ represents element-wise multiplication) is calculated, resulting in [0.3, 0.55, 0.13, 0.03, 0.02]. V_adj is normalized so that its element-wise sum is 1, resulting in V_norm. The new class code corresponding to the maximum probability value in V_norm is found. If the new code is different from the original dominant class code in V_orig, the class of that pixel in time phase T is updated with the new code. If an update occurs, this update may affect the class determination of its spatial neighboring pixels in times phases T-1, T, and T+1, so the information of these affected neighboring pixels needs to be added back to the tail of the processing queue. The submodule continuously takes tasks from the queue, calculates, judges, and updates until the queue is empty, finally obtaining an optimized class data for all pixels after temporal smoothing and consistency correction.

[0027] The optimization result integration submodule is responsible for reconstructing the final results. Based on the optimized category data, it merges adjacent pixels with the same category code in space to generate a temporally consistent land cover patch coding sequence with a unique identifier, spatial outline, and temporal category coding sequence. Simultaneously, using this optimized data, a change detection and regional aggregation process similar to that in Example 2 is re-executed to generate a higher-quality optimized land cover change polygon set. This embodiment, through a closed loop of "screening – iterative optimization – integration," transforms geographical common sense (that land cover types have continuity on short-term time scales) into computable optimization criteria, significantly improving the reliability of monitoring results in the time dimension. Example

[0028] Based on Embodiment 1 and Embodiment 3, the credible change identification module includes a change confidence calculation submodule and a confidence grading submodule.

[0029] This embodiment details the automatic assessment and grading process for the credibility of change results. The change confidence calculation submodule traverses each polygon in the optimized land cover change polygon set. For a given polygon, the submodule first obtains a list of coordinates of all pixels it covers, as well as the specific time phase of the land cover change recorded in the polygon's attributes (e.g., from 2023 to 2024). Next, based on these pixel coordinates and the time phase of the change (2024), the submodule backtracks to query the multi-temporal pixel-level category probability matrix, reads the classification probability vector of each pixel in the 2024 time phase, and precisely extracts the probability component value from the pixel probability vector that corresponds to the "post-change category code" recorded by the polygon. Assuming a polygon contains 50 pixels and its recorded "post-change category" is "woodland" (corresponding to code 1), the submodule will extract the probability values ​​corresponding to code 1 (woodland) from the probability vectors of each of these 50 pixels, for a total of 50. Subsequently, the submodule calculates the arithmetic mean of these 50 probability values. This average value is the polygon average change confidence level, which represents the overall classification confidence level of the changed area.

[0030] The credibility grading submodule receives the average confidence level of all polygon changes. It presets a high-confidence threshold for judgment. The specific value of this threshold can be determined based on historical data verification results, the accuracy requirements of the application scenario, or experimental analysis; for example, it can be set to 0.75. The submodule compares the average confidence level of each polygon with this threshold. As an example, if the threshold is 0.75, polygons with an average confidence level greater than 0.75 are judged as high-confidence changes and included in the high-confidence change set; polygons with an average confidence level less than or equal to 0.75 have credibility that needs further verification and are included in the change-to-be-verified set. Through this automatic grading operation, the system transforms the original, unverified list of change results into two subsets with different priorities. This allows business personnel to prioritize and analyze changes in the high-confidence set, which have a very high probability of actually occurring; while for the change-to-be-verified set, targeted manual review or supplementary investigation can be arranged, thereby greatly optimizing the allocation of human resources and improving the efficiency of the entire monitoring process and the practicality of the output results. Example

[0031] Based on Embodiment 1, the multi-temporal data loading module includes a data coupling loading submodule and a pixel-level probability generation submodule.

[0032] This embodiment illustrates the specific implementation of the "data-model coupling driven" approach in the initial stage of the system. The data coupling loading submodule is responsible for the unified scheduling and preparation of data. It retrieves multi-temporal high-resolution remote sensing image data of the target area from the storage system (e.g., GeoTIFF format files containing multiple spectral bands such as red, green, blue, and near-infrared). Simultaneously, it calls the deep neural network model file that has been trained offline and persistently stored. This file contains parameters such as the weight matrices and bias vectors of all layers of the network. The submodule synchronously loads the image data (pixel value matrix) and model parameter data into the efficient access area of ​​the graphics processing unit (GPU)'s video memory or main memory, forming a loaded set of image data and model parameters. This "coupled loading" ensures that subsequent calculations can access the required data and model at high speed and directly, which is the foundation for efficient processing.

[0033] The pixel-level probabilistic generation submodule is the core of the model inference. Based on a loaded set, it performs forward propagation computation on each pixel in the image in parallel by the GPU. Taking a convolutional neural network-based model as an example, for each pixel and its surrounding neighborhood image patch, the computation unit performs a series of predefined operations: first, convolution is performed, using the loaded convolution kernel weights to extract features from multi-band spectral values; then, a non-linear activation function (such as ReLU) is applied; pooling operations may also be performed to reduce spatial dimensionality and enhance feature robustness. These operations are stacked layer by layer, and finally, through a fully connected layer, the learned high-level features are mapped to the score of each predefined land cover category. Finally, these scores are normalized by the Softmax function and converted into a probability distribution vector. For example, for a certain pixel, the output vector might be [0.02, 0.95, 0.01, 0.01, 0.01], indicating that the pixel has a 95% probability of belonging to cultivated land and a 2% probability of belonging to forest land, etc. After performing the above calculations on all pixels of the entire image, the system organizes the coordinates, temporal information, and corresponding category probability vectors of each pixel, thus generating the multi-temporal pixel-level category probability matrix upon which all subsequent modules rely for operation. This embodiment demonstrates how to closely integrate a pre-trained AI model with remote sensing image data to achieve automated and intelligent conversion from raw spectral values ​​to advanced semantics (land cover category probability).

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A remote sensing image land cover monitoring system based on data-model coupling, characterized in that, This includes the following processing modules connected in sequence: The multi-temporal data loading module calls the neural network parameters, and the calculation unit performs hierarchical transformation on the pixel spectral values, outputs the probability of its classification into each category, and generates a multi-temporal pixel-level category probability matrix. The change region extraction module, based on a multi-temporal pixel-level category probability matrix, compares the maximum probability category encoding of pixels in adjacent temporal phases, records pixels with different encodings, and aggregates spatially adjacent pixels with the same change category to generate a set of land cover change polygons. The temporal consistency optimization module calls the multi-temporal pixel-level category probability matrix and the set of land cover change polygons to extract the temporal category and probability of pixels within the polygons. It determines whether the probability of a single point change is lower than a set threshold or whether the number of overall state changes exceeds a set threshold. The coordinates of suspicious pixels that meet the judgment conditions and their target temporal information are added to the queue to be processed. The original probability vectors of the pixels in the queue are weighted, adjusted and normalized according to the temporal categories of the pixels before and after. The pixel categories are calculated and updated in a loop until the queue is empty, resulting in a temporally consistent land cover patch coding sequence and an optimized set of land cover change polygons. The credible change identification module calculates the average probability value of pixels within each polygon during the change phase for the optimized set of land cover change polygons, compares this average value with a preset high confidence threshold, and classifies the polygons into a high confidence change set and a change set to be verified based on the results. The structured monitoring product generation module calls the time-consistent land cover patch coding sequence, high-confidence change set, and change set to be verified, and encapsulates the data according to the format for the spatial, category, and confidence identification attributes of the patches and polygons, generating a land cover monitoring product set with confidence level labels.

2. The remote sensing image land cover monitoring system based on data-model coupling as described in claim 1, characterized in that: The multi-temporal pixel-level category probability matrix is ​​specifically a type of structured data, including the spatial coordinates of each pixel, the corresponding image acquisition temporal identifier, and a probability value vector representing that the pixel belongs to each predefined land cover category. The land cover change polygon set is specifically a set of geographic features, where each feature includes a geometric outline defined by a set of continuous vertex coordinates, the dominant land cover category code before the change represented by the polygon, and the dominant land cover category code after the change. The temporally consistent land cover patch coding sequence is specifically a set of temporal geographic objects, including the unique identifier of each patch, its spatial outline in different temporal phases, and the optimized dominant land cover category coding sequence arranged in chronological order. The optimized land cover change polygon set is specifically an optimized subset of the land cover change polygon set, wherein each optimized change polygon includes an optimized geometric profile, an optimized pre-change category code, an optimized post-change category code, and a state identifier used to identify that it has undergone temporal consistency optimization. Both the high-confidence change set and the change set to be verified refer to subsets divided from the optimized land cover change polygon set according to specific rules. Each subset includes a confidence level identifier and a change event attribute table corresponding to all change polygons in the subset. The attribute table includes change area, change phase, and average change confidence. The land cover monitoring product set with confidence level identifiers is specifically the final system output data package, which includes data corresponding to the time-consistent land cover patch coding sequence, vector layers of high-confidence change patches, vector layers of change patches to be verified, and product metadata files describing the data source, processing time, and coordinate system.

3. The remote sensing image land cover monitoring system based on data-model coupling as described in claim 1, characterized in that, The change region extraction module includes a change pixel recording submodule and a change region aggregation submodule; The variable pixel recording submodule, based on the multi-temporal pixel-level category probability matrix, reads the probability value array of pixels in adjacent temporal phases and finds the category code corresponding to the highest probability, compares whether the category codes of two temporal phases are equal, and records the spatial coordinates and the category codes of the preceding and following temporal phases for pixels with unequal category codes, and generates a variable pixel recording table. The change region aggregation submodule calls the change pixel record table, searches for other records with the same preceding and following category codes in the eight neighboring regions based on the record coordinates, and merges the record pixels that meet the conditions into continuous polygons in space to generate a set of land cover change polygons.

4. The remote sensing image land cover monitoring system based on data-model coupling as described in claim 1, characterized in that: The temporal consistency optimization module includes a suspicious pixel screening submodule, a pixel category iterative optimization submodule, and an optimization result integration submodule; The suspicious pixel screening submodule calls the multi-temporal pixel-level category probability matrix and the land cover change polygon set. For pixels within the polygon, it determines whether the probability of change at a single point is lower than the lower limit of classification confidence or whether the number of changes in the overall state of the polygon exceeds the state fluctuation threshold. The pixel coordinates and target temporal information that meet the conditions are added to the queue to be processed. The pixel category iterative optimization submodule, based on the queue to be processed and the multi-temporal pixel-level category probability matrix, retrieves pixel information from the queue, performs weighted adjustment and normalization on the original probability vector according to its preceding and following temporal categories, calculates its optimized category, updates it if the category changes, and adds the affected pixel information to the tail of the queue, loops until the queue is empty, and obtains the optimized category data of all pixels. The optimization result integration submodule calls the optimized category data of all pixels, aggregates spatially adjacent pixels with the same category into patches, associates each temporal category code with the spatial contour, generates a temporally consistent land cover patch code sequence, and re-detects and aggregates change areas based on this to generate an optimized land cover change polygon set.

5. The remote sensing image land cover monitoring system based on data-model coupling as described in claim 1, characterized in that: The credible change identification module includes a change confidence calculation submodule and a confidence grading submodule; The change confidence calculation submodule, for the optimized set of land cover change polygons, traverses each polygon in the set, extracts all pixel coordinates covered by it, reads the corresponding dominant class probability value from the multi-temporal pixel-level class probability matrix based on these coordinates and the time phase of the change, calculates the arithmetic mean of these read probability values, and obtains the average change confidence of the polygon. The confidence level submodule calls the average change confidence of polygons, compares it with the preset high confidence threshold value, classifies polygons with an average change confidence greater than the threshold into a new set, and classifies polygons with an average change confidence less than or equal to the threshold into another new set, generating a high confidence change set and a change set to be verified.

6. The remote sensing image land cover monitoring system based on data-model coupling as described in claim 4, characterized in that: In the suspicious pixel identification submodule, the determination of whether the probability value of a pixel in an adjacent time phase is less than the lower limit of classification confidence is specifically performed as follows: For a marked pixel, retrieve the two specific time phases T_a and T_b involved in the change of its category, read the dominant category probability value P_a of the pixel in time phase T_a and the dominant category probability value P_b of the pixel in time phase T_b, and determine whether P_a is less than the lower limit of classification confidence or whether P_b is less than the lower limit of classification confidence.

7. The remote sensing image land cover monitoring system based on data-model coupling as described in claim 4, characterized in that: In the iterative optimization submodule, when constructing the weight adjustment vector, the fixed increment of the element values ​​with indices equal to L_prev and L_next is a constant greater than zero.

8. The remote sensing image land cover monitoring system based on data-model coupling as described in claim 1, characterized in that: The structured monitoring product generation module includes an attribute association encapsulation submodule and a product packaging output submodule; The attribute association encapsulation submodule calls the time-consistent land cover patch coding sequence, high-confidence change set, and change set to be verified to generate a set of geographic features with attributes for the spatial geometry, category coding, and confidence identification attributes associated with patches and polygons. The product packaging and output submodule, based on a set of geographic features with attributes, encodes the geographic features according to the specified spatial data format and creates a metadata file containing data source, temporal phase and coordinate system information. After packaging, it generates a land cover monitoring product set with a reliability rating identifier.

9. A remote sensing image land cover monitoring system based on data-model coupling as described in claim 1, characterized in that: The multi-temporal data loading module includes a data coupling loading submodule and a pixel-level probability generation submodule; The data coupling loading submodule acquires multi-temporal high-resolution remote sensing image data of the target area, calls the stored deep neural network model weight parameter file, and loads the image data and model weight parameters into the memory of the graphics processing unit simultaneously, generating a set of loaded image data and model parameters. The pixel-level probability generation submodule, based on the loaded image data and model parameter set, performs continuous convolution and nonlinear activation operations on the multi-band spectral values ​​corresponding to each pixel in the image. Finally, it outputs a numerical array representing the probability of the pixel belonging to each predefined land cover category. After completing the calculation for all pixels, it generates a multi-temporal pixel-level category probability matrix.

Images