Remote sensing image adaptive identification method and system for territorial space planning

Abstract

The invention relates to the technical field of remote sensing image processing, and discloses a remote sensing image adaptive identification method and system for territorial space planning, and the method comprises the steps: obtaining a multi-source remote sensing image data set of a research region, feature extraction, cloud detection, quality evaluation and adaptive preprocessing are carried out; carrying out prototype network coding, calculating a category prototype and probability, and supporting fine tuning of a set; carrying out multi-scale cavity convolution and category scale attention fusion; evaluating the adaptability score of the comprehensive fusion feature map set, and carrying out weighted fusion, classification and normalization; change detection is carried out, stable and change regions are segmented, and time sequence context features are extracted and constrained optimization is carried out; entropy is fused, a boundary is decided, uncertainty is estimated, and weighted fusion is carried out according to a change area; conditional random field optimization, confidence level grading and connected domain identification are carried out; the automation level, the adaptive capacity and the recognition reliability of remote sensing monitoring of territorial space planning are improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and more specifically, to an adaptive recognition method and system for remote sensing images in land spatial planning. Background Technology

[0002] Territorial spatial planning is a crucial institutional arrangement for coordinating the development and protection of territorial space, optimizing its spatial pattern, and enhancing its governance capabilities. It is vital for sustainable economic and social development and the construction of ecological civilization. Remote sensing technology, with its advantages of wide coverage, rapid updates, and objective accuracy, has become a core technical means for conducting surveys on the current status of territorial spatial planning, as well as for planning formulation and implementation supervision. Remote sensing imagery can quickly obtain information on current land use, monitor illegal and irregular land use, and assess the effectiveness of planning implementation, providing a scientific basis for territorial spatial management.

[0003] In practical applications, especially in mountainous counties with complex terrain, the acquisition and processing of remote sensing imagery faces numerous challenges. Due to terrain obstruction and climatic conditions, high-quality remote sensing images in mountainous areas are limited in temporal phase, with severe cloud and fog obstruction. Images acquired at different times and from different sensors exhibit significant differences in spatial resolution, spectral characteristics, and radiometric quality, resulting in inconsistent image quality. Territorial spatial planning involves a complex and diverse range of land features, including urban construction land, rural settlements, various types of agricultural land, forests and grasslands, water bodies, and more than ten other land categories. Continuous dynamic monitoring of these land categories is necessary to promptly detect changes in land features. Traditional remote sensing image recognition methods are ill-suited to these complex and ever-changing application requirements, primarily exhibiting the following technical problems: lack of targeted image quality assessment leads to inappropriate preprocessing strategies; insufficient generalization ability of deep learning models necessitates extensive retraining with labeled samples; rigid multi-source data fusion strategies cannot be adaptively adjusted; lack of consistency constraints in temporal monitoring leads to category jumps; and the lack of uncertainty assessment in recognition results impacts the efficiency of manual verification.

[0004] While existing technologies have made some progress in remote sensing image classification and change detection, research on adaptive identification methods for the specific application scenario of land and space planning is still insufficient. There is a need to develop new identification technologies with capabilities such as adaptive image quality assessment, rapid model transfer, multi-source intelligent fusion, temporal dynamic constraints, and uncertainty quantification to meet the urgent needs of land and space planning for automated, intelligent, and precise remote sensing monitoring. Summary of the Invention

[0005] This invention provides an adaptive recognition method and system for remote sensing images in land spatial planning, which solves the technical problems in related technologies where the lack of targeted image quality assessment leads to inappropriate preprocessing strategies and the rigidity of multi-source data fusion strategies that cannot be adaptively adjusted.

[0006] This invention provides an adaptive recognition method for remote sensing images in land spatial planning, comprising: A multi-source remote sensing image dataset of the study area was acquired, and feature extraction, cloud detection and quality assessment, and adaptive preprocessing were performed to obtain an optimized image dataset. Based on the optimized image dataset, prototype network encoding, category prototype and probability calculation, and support set fine-tuning are performed to obtain feature maps adapted to the target domain. Based on the feature map adapted to the target domain, multi-scale dilated convolution and category-scale attention are fused to obtain a comprehensive fused feature map set. The suitability score of the integrated feature map set is evaluated and weighted, fused, classified and normalized to obtain preliminary classification results; Based on the preliminary classification results, change detection and segmentation of stable and change regions are performed, temporal context features are extracted and constraint optimization is carried out to obtain the temporally optimized classification results. Based on the classification results optimized by time series, the information entropy and uncertainty of the decision boundary are calculated, and the confidence score is obtained by weighted fusion according to the change region. Based on the confidence score, conditional random field optimization, confidence leveling and connected component identification are performed to generate a priority-ranked verification task list.

[0007] In a preferred embodiment, the optimized image dataset specifically includes: A multi-scale convolutional neural network is used to extract features in parallel across three feature channels: spectral, texture, and spatial structure, to obtain a multi-dimensional feature representation. Based on multidimensional feature representation and digital elevation model, a dedicated deep learning model for cloud detection is used to identify cloud-covered areas and calculate cloud coverage. A no-reference image quality assessment network is used to extract basic quality vectors, and a task relevance assessment module is used to calculate task quality scores for each land cover category. A multidimensional quality assessment vector is constructed based on the basic quality vector, task quality score and cloud coverage. The preprocessing parameter vector is determined by selecting the network through an adaptive preprocessing strategy. Radiometric correction, contrast enhancement, denoising, sharpening, and dehazing are performed based on the preprocessing parameter vector to obtain an optimized image dataset.

[0008] In a preferred embodiment, obtaining the feature map adapted to the target domain specifically includes: Deep feature encoding is performed on the optimized image dataset using a pre-trained prototype network encoder to obtain embedded feature maps. Based on the target domain support set, the mean vector of the feature vector of the support samples for each land cover category is calculated as the category prototype vector, forming a prototype set; Calculate the Euclidean distance between the feature vector at each spatial location in the embedded feature map and the prototype vector of each category, convert the distance into a similarity score and normalize it to obtain the category probability map; Based on the target domain support set, the gradient descent optimization algorithm is used to fine-tune the encoder parameters and adjust the parameters of the last preset number of layers of the encoder. The image features are re-extracted using the adapted encoder to obtain a feature map adapted to the target domain.

[0009] In a preferred embodiment, obtaining the integrated fusion feature map set specifically includes: A multi-branch dilated convolutional network is used to extract feature representations at different receptive field scales, with four parallel dilated convolutional branches. Based on the prior knowledge base of land cover scale, an association matrix between land cover categories and spatial scale is constructed. The association matrix represents the correlation strength between each land cover category and each void ratio. Scale response maps are generated from feature maps at each scale. The scale weight vector of each category at each spatial location is calculated by combining the category scale correlation matrix and the scale response map. For each category, the feature maps at the four scales are weighted and fused according to their corresponding scale weights to obtain a category-specific fused feature map; By concatenating channels and performing convolutional dimensionality reduction on the fused feature maps of all categories, a comprehensive fused feature map set is obtained.

[0010] In a preferred embodiment, obtaining the preliminary classification result specifically includes: The input consists of fused feature maps and category prototype vectors from the data source. A data source category fit evaluation network is used to quantify the ability of each data source to identify different land cover categories and construct a fit matrix. Based on the adaptability matrix and the multidimensional quality assessment vectors of each data source, the basic quality score of the data source is multiplied by its adaptability score for each category and normalized to construct a category-related fusion weight matrix. For each category, the fusion feature maps from each data source are weighted and fused according to the corresponding fusion weights to obtain a category-specific multi-source fusion feature map; A dedicated binary classifier is built for each category. The multi-source fusion feature map is classified to obtain a category probability map. The probability maps of all categories are normalized and the maximum value is taken to obtain the preliminary classification result.

[0011] In a preferred embodiment, obtaining the time-optimized classification result specifically includes: A twin change detection network is used to identify significant change regions between adjacent time phases, calculate the difference features of the feature maps of the two time phases, and output a change probability map; An adaptive threshold segmentation method is used to divide the image region into changing regions and stable regions, generating changing region masks and stable region masks; The category probability maps of each time phase are multiplied with the stable region mask, and a bidirectional gated recurrent unit network is used to extract the temporal context features. The correlation strength between the current time phase and the adjacent time phases is calculated by using an attention mechanism. Different constraint strength coefficients are set according to the region type. Strong temporal consistency constraints are applied to stable regions, and weak temporal constraints are applied to changing regions, so as to obtain the classification results after temporal optimization.

[0012] In a preferred embodiment, obtaining the confidence score specifically includes: The information entropy of the category probability vector for each spatial location is calculated to obtain an entropy-based uncertainty map; The difference between the highest probability and the second highest probability is calculated and converted into an uncertainty measure to obtain an uncertainty map based on the decision boundary; A lightweight deep ensemble architecture with a shared feature extractor and five classification heads is adopted to calculate the variance of the prediction results of the five classification heads and obtain the model uncertainty. The three uncertainty measures are normalized and weighted summed. The fusion uncertainty is adjusted according to the region-dependent mask based on the change region, and the adjusted uncertainty is converted into a confidence score.

[0013] In a preferred embodiment, the generation of priority-ordered verification task sheets specifically includes: A conditional random field model is used to perform spatial consistency post-processing on the classification results. An energy function containing univariate and binary potential functions is constructed, and the spatially optimized classification results are obtained by minimizing the energy function. Set high confidence thresholds and low confidence thresholds to divide the recognition results into high confidence region, medium confidence region and low confidence region; For regions with low confidence, perform connected component analysis, select connected components with an area greater than a preset minimum threshold as verification units, and sort them from low to high according to average confidence. The classification results are converted into vector format using a raster-to-vector algorithm to generate a priority-ranked checklist.

[0014] In a preferred embodiment, the adaptive preprocessing strategy selection network includes: The denoising parameters are automatically adjusted based on the signal-to-noise ratio score, and strong denoising parameters are selected for images with low signal-to-noise ratio. Adjust the contrast enhancement parameters according to the contrast score, and select strong contrast enhancement parameters for images with insufficient contrast. Whether to enable the dehazing algorithm is determined based on the cloud coverage level; for images with high cloud coverage, dehazing processing is automatically enabled. For high-quality images, use lightweight processing to avoid over-enhancement, and for low-quality images, use enhanced processing to improve data usability.

[0015] This invention provides an adaptive recognition system for remote sensing images in land spatial planning, used to execute the aforementioned adaptive recognition method for remote sensing images in land spatial planning, comprising: The data acquisition and preprocessing module is used to acquire multi-source remote sensing image datasets of the study area, perform feature extraction, cloud detection and quality assessment, and adaptive preprocessing to obtain optimized image datasets. The prototype network adaptive module, based on the optimized image dataset, performs prototype network encoding, calculates category prototypes and probabilities, and fine-tunes the support set to obtain feature maps adapted to the target domain. The multi-scale feature fusion module performs multi-scale dilated convolution and category-scale attention fusion based on the feature map adapted to the target domain to obtain a comprehensive fusion feature map set; The multi-source data fusion and classification module is used to evaluate the suitability score of the integrated fusion feature map set and perform weighted fusion, classification and normalization to obtain preliminary classification results; The temporal consistency optimization module, based on the preliminary classification results, detects changes and segments stable and changing regions, extracts temporal context features and performs constraint optimization to obtain the temporally optimized classification results. The uncertainty assessment module calculates the information entropy and uncertainty of the decision boundary based on the time-series optimized classification results, and obtains the confidence score by weighted fusion according to the change region. The post-processing and result generation module performs conditional random field optimization, confidence leveling, and connected component identification based on confidence scores, and generates priority-ranked verification task sheets.

[0016] The beneficial effects of this invention are as follows: By constructing a multi-task-oriented image quality assessment mechanism, not only is the physical quality of the images evaluated, but also their adaptability to different land cover category recognition tasks. Personalized preprocessing strategies are automatically selected based on multi-dimensional quality vectors, enabling the system to adaptively process input images of varying qualities. Through a prototype network-based meta-learning method, the model achieves rapid adaptation under conditions of minimal samples. Only 5 to 10 samples per class are needed for the pre-trained model to adapt to new research areas, eliminating the need to collect large amounts of labeled data and retrain the model for each area, thus reducing application costs and time overhead.

[0017] The category-aware multi-source data fusion strategy proposed in this invention breaks through the limitations of traditional fixed-weight fusion. By establishing a data source-category adaptability matrix, the fusion weights are dynamically adjusted based on the recognition advantages of different data sources for different land cover categories, giving full play to the complementary advantages of each data source and significantly improving the fusion effect. By introducing change detection-guided temporal consistency constraints, strong constraints are applied in stable regions to ensure temporal continuity, while weak constraints are applied in changing regions to allow for real changes, effectively solving the problem of unreasonable category jumps of the same land parcel in adjacent time phases in traditional methods. Attached Figure Description

[0018] Figure 1 This is a flowchart of a remote sensing image adaptive recognition method for land spatial planning according to the present invention; Figure 2 This is a block diagram of a remote sensing image adaptive recognition system for land spatial planning according to the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses an adaptive recognition method for remote sensing images in land spatial planning, such as... Figure 1 As shown, it includes: Step 1: Obtain a multi-source remote sensing image dataset of the study area, perform feature extraction, cloud detection and quality assessment, and adaptive preprocessing to obtain an optimized image dataset; Step 1.1: Obtain the dataset of multi-source, multi-temporal remote sensing images to be identified and the set of target ground cover categories; Based on the requirements of land and space planning monitoring, a multi-source remote sensing image dataset of the study area was acquired. This dataset includes remote sensing data from various sources, such as Gaofen-2 optical satellite imagery, Sentinel-2 multispectral imagery, commercial Gaofen satellite imagery, and UAV orthophotos. The spatial resolution of each image ranges from 0.5 meters to 30 meters, and the number of spectral bands varies from 3 to 13. The imaging time spans multiple time phases across the four seasons of the year. Simultaneously, a set of target identification land cover categories was acquired, including 13 standard land use categories for land and space planning: urban construction land, rural residential areas, industrial and mining land, transportation land, paddy fields, dry land, orchards, forest land, shrubland, grassland, river surfaces, reservoirs and ponds, and unused land. Furthermore, digital elevation model (DEM) data of the study area was acquired to assist in cloud and fog identification and terrain analysis.

[0021] Step 1.2: Extract multidimensional feature representations of the image; Based on the input remote sensing imagery, a multi-scale convolutional neural network is employed to extract features in parallel across three feature channels: spectral, texture, and spatial structure. In the spectral feature channel, 1×1 convolutional kernels are used to perform linear combinations and nonlinear transformations on each band to extract features reflecting the spectral response characteristics of ground features. In the texture feature channel, 3×3 and 5×5 convolutional kernels are used to extract local texture patterns and edge information. In the spatial structure feature channel, 7×7 and 9×9 large-scale convolutional kernels are used to capture the morphological and spatial layout features of ground features. Each of the three feature extraction networks contains three convolutional layers, followed by batch normalization and a ReLU activation function. The features extracted from the three channels are concatenated along the channel dimension to obtain a multi-dimensional feature representation, comprising spectral features, texture features, and spatial structure features. Its spatial dimension is the image height multiplied by its width, and its feature dimension is the total number of feature channels across the three channels.

[0022] Step 1.3: Identify areas obscured by clouds and fog and calculate cloud coverage. Based on multidimensional feature representation and a digital elevation model (DEM), a dedicated deep learning model for cloud detection is used to identify cloud and fog-occupied areas in imagery. The cloud detection model employs an improved U-Net architecture, where the encoder extracts multi-scale features from the feature representation, and the decoder progressively restores spatial resolution and outputs pixel-level cloud and fog probability maps. DEM data is introduced as an additional input channel into the network, and by learning the correlation between elevation and cloud and fog distribution, the accuracy of cloud and fog identification in mountainous environments is improved. The model is pre-trained on a large-scale cloud and fog annotation dataset, learning spectral and texture difference patterns for different types of clouds, fog, snow, and bright surfaces.

[0023] A cloud coverage probability map is obtained by forward inference of multidimensional feature representation and DEM data using a cloud detection function. The value of each pixel in the probability map represents the probability that the location is obscured by clouds or fog. A threshold segmentation operation is performed on the probability map, marking pixels with a probability value higher than the threshold as 1 to indicate cloud or fog obscuration, and pixels with a probability value lower than the threshold as 0 to indicate no clouds, resulting in a binary cloud coverage mask with a spatial size equal to the height multiplied by the width of the image. The number of all cloud or fog pixels marked as 1 in the mask is counted, and this number is divided by the total number of pixels in the image, i.e., the height multiplied by the width, to calculate the cloud coverage rate. This rate reflects the proportion of the image area obscured by clouds or fog.

[0024] Step 1.4: Assess the basic quality and mission-related quality of the imagery; Based on multidimensional feature representation and cloud cover masking, a no-reference image quality assessment network is employed to extract basic image quality metrics in non-cloud areas. The quality assessment network uses a lightweight convolutional neural network structure, taking feature representations as input and outputting a quality vector containing signal-to-noise ratio (SNR) scores, contrast scores, and sharpness scores. The SNR score estimates the noise level by analyzing the variance of smooth regions of the features; the contrast score assesses the dynamic range by calculating the distribution range of the feature histogram; and the sharpness score assesses image sharpness by extracting the intensity of high-frequency edge information. The quality assessment network performs calculations in non-cloud areas marked by the cloud cover mask to avoid interference from cloud and fog regions in the quality assessment.

[0025] The multidimensional feature representation and cloud cover mask are input into the quality evaluation function, and the basic quality vector is obtained through network inference. This vector contains three components, which correspond to the signal-to-noise ratio score, contrast score, and sharpness score, respectively. The value of each component is between 0 and 1, and the larger the value, the better the quality of that item.

[0026] Based on the target recognition land cover category set and image features, a task relevance assessment module is used to calculate the applicability of the image to different land cover categories. The task relevance assessment is built upon a land cover category prototype library, which stores typical representation vectors of each land cover category in the feature space. These prototype vectors are obtained by calculating the mean of features of similar samples on a large number of labeled samples. For the current image, its global feature statistics are extracted as the overall feature representation of the image. The cosine similarity between this global feature and the prototype vectors of each category is calculated. Specifically, the global feature vector is multiplied by the prototype vector of a certain category, and then divided by the product of the norms of the two vectors to obtain a normalized similarity value. The above similarity calculation is performed on each of the 13 land cover categories to obtain a task quality score for each land cover category. The score ranges from 0 to 1. A high task quality score indicates that the feature distribution of the current image matches the typical features of that category well, making it suitable for the recognition of that category.

[0027] Step 1.5: Construct a multidimensional quality assessment vector and select an adaptive preprocessing strategy; Based on the base quality vector, task quality scores for each category, and cloud coverage, a comprehensive multi-dimensional quality assessment vector is constructed using vector concatenation. The three components of the base quality vector—signal-to-noise ratio score, contrast score, and sharpness score—are arranged sequentially; task quality scores for 13 land cover categories are concatenated; and cloud coverage is concatenated to form a comprehensive quality assessment vector containing 17 elements. This quality assessment vector comprehensively characterizes the image's physical quality, information quality, and task suitability.

[0028] Based on a multidimensional quality assessment vector, an adaptive preprocessing strategy selection network is employed to determine the optimal preprocessing scheme. The strategy selection network uses a multilayer perceptron structure, taking the quality assessment vector as input and outputting a preprocessing parameter vector. This vector contains five parameters corresponding to the settings of radiometric correction intensity, contrast enhancement parameters, denoising intensity, sharpening degree, and dehazing processing method. The network is trained on a large amount of paired data of images with different quality levels and their optimal preprocessing parameters, learning the mapping relationship between the quality vector and the preprocessing parameters.

[0029] The specific adaptive logic is as follows: the quality assessment vector is input into the strategy selection network, and the network automatically adjusts the denoising parameters according to the signal-to-noise ratio score, selecting strong denoising parameters for images with low signal-to-noise ratio; the contrast enhancement parameters are adjusted according to the contrast score, selecting strong contrast enhancement parameters for images with insufficient contrast; whether to enable the dehazing algorithm is determined according to the cloud coverage rate, and dehazing processing is automatically enabled for images with high cloud coverage; a personalized preprocessing strategy parameter set is obtained through network inference, and this parameter set is customized for the quality characteristics of the current image.

[0030] Step 1.6: Perform adaptive preprocessing to obtain the optimized image dataset; Based on the preprocessing strategy parameter set and the original image, a parametric preprocessing pipeline sequentially executes various preprocessing operations. Atmospheric correction is performed according to the radiometric correction parameters, converting apparent reflectance to surface reflectance to correct for the effects of atmospheric scattering and absorption. Adaptive histogram equalization is used to enhance image contrast based on the contrast enhancement parameters, achieving local contrast enhancement through block processing and bilinear interpolation. Nonlocal mean filtering or bilateral filtering algorithms are used to suppress image noise based on the denoising parameters, smoothing noise while preserving edge details. Laplacian sharpening or unsharpening masks are used to enhance edge and texture information based on the sharpening parameters. For high-cloud-covered images, temporal replacement or interpolation restoration strategies are enabled based on the de-clouding parameters, using cloud-free images from adjacent time phases to fill in cloud-covered areas. Each stage of the preprocessing pipeline adaptively adjusts the processing intensity according to the parameter strength, using lightweight processing to avoid over-enhancement for high-quality images and enhanced processing to improve data usability for low-quality images, resulting in an optimized image dataset.

[0031] Based on the optimized image dataset and cloud cover mask, the availability mask is updated. Areas that cannot be restored after dehazing are marked as unavailable, while successfully restored cloud-covered areas and originally cloudless areas are marked as available. The availability mask is used to guide the model to identify only reliable areas in subsequent processing.

[0032] In some embodiments, due to the complex terrain of mountainous areas leading to severe shadow effects, a terrain shadow correction method can be used to further process the optimized image. Based on the digital elevation model (DEM) and the solar altitude angle and azimuth angle at the time of image imaging, the slope, aspect, and shadow mask at each location on the ground surface are calculated. The C-correction model or the Minnaert correction model is used to compensate for the radiance values ​​in the shadowed areas. The specific correction process is as follows: calculate the product of the cosine of the solar zenith angle and the cosine of the surface normal zenith angle, add the terrain correction coefficient C to obtain the numerator; calculate the cosine of the incident angle and add the terrain correction coefficient C to obtain the denominator; divide the numerator by the denominator to obtain the correction factor; multiply the observed radiance value by the correction factor to obtain the corrected radiance value. The observed radiance value is the actual measured radiance intensity in the image, the solar zenith angle represents the angle between the sunlight and the vertical direction of the ground surface, the surface normal zenith angle represents the angle between the surface normal and the vertical direction, the incident angle represents the angle between the sunlight and the surface normal, and the terrain correction coefficient C is determined according to the surface type and terrain features. The purpose of this correction method is to reduce the impact of terrain shadows on the spectral characteristics of ground features and improve the accuracy of shadowed area identification.

[0033] Step 2: Based on the optimized image dataset, perform prototype network encoding, calculate category prototypes and probabilities, and fine-tune the support set to obtain a feature map adapted to the target domain. Step 2.1: Extract image embedding features using a pre-trained prototype network encoder; Based on an optimized image dataset and a pre-trained prototype network model, a feature extraction encoder is employed to perform deep feature encoding on the images. The encoder uses a ResNet50 or EfficientNet backbone network structure, containing multiple residual blocks or MBConv modules, and extracts multi-level feature representations of the images through layer-by-layer convolution, pooling, and non-linear activation. The encoder is pre-trained on a large-scale remote sensing image dataset, learning general land cover recognition knowledge, including the spectral patterns, texture features, and spatial structures of different land covers.

[0034] The optimized image dataset is input into the encoder, where it undergoes convolution, pooling, and nonlinear activation transformations in each layer of the encoder network. Features are extracted layer by layer through forward propagation. A high-dimensional embedding feature map is obtained at the encoder's output layer. The spatial dimensions of this feature map are the downsampled height and width, and the feature dimension is the number of feature channels. The embedding feature map transforms the spectral information of the original image into a high-level semantic representation, providing a foundation for subsequent prototype matching and classification.

[0035] Step 2.2: Calculate the prototype vectors for each category based on the target domain support set; Based on the target domain support set, a prototype computation strategy is used to obtain the prototype representation of each land cover category in the feature space. The target domain support set is a small number of labeled samples collected from the current study area, typically containing only 5 to 10 samples per category, with the total number of samples K usually ranging from 65 to 130.

[0036] For each sample in the support set, its embedded features are extracted by the encoder to obtain the vector representation of the sample in the high-dimensional feature space. For each land cover category, the feature vectors of all support samples of that category are collected, these feature vectors are summed, and then divided by the number of samples in that category to calculate the mean vector of all sample features of that category. This mean vector is the prototype vector of that category. The prototype vector represents the center position of the category in the feature space and reflects the typical feature pattern of the category. The above prototype calculation process is performed on each of the 13 land cover categories to obtain the prototype vectors of the 13 categories. These prototype vectors are then aggregated to form a prototype set.

[0037] Step 2.3: Calculate the distance between the query feature and the prototype and obtain the category probability; Based on the prototype set and image embedding feature map, a distance metric function is used to calculate the similarity between the feature vector of each spatial location in the feature map and the prototypes of each category.

[0038] For a feature vector at a spatial location in the feature map, calculate its Euclidean distance with each class prototype vector. Specifically, subtract the feature vector at that location element-wise from the class prototype vector to obtain a difference vector. Calculate the L2 norm of the difference vector, which is the square root of the sum of squares of all elements, to obtain the Euclidean distance. The smaller the distance, the more similar the feature at that location is to the class prototype, and the greater the probability that it belongs to that class. Convert the distance to a similarity score using the exponential form of the negative distance, i.e., calculate the negative power of the natural constant e to obtain the similarity score for that location with respect to that class. The similarity scores of 13 categories are subjected to softmax normalization. Specifically, the sum of the similarity scores of all categories is calculated as the normalization denominator, and the similarity score of each category is divided by the sum to obtain the normalized probability value. This ensures that the sum of the probabilities of all categories is 1, thus obtaining the probability distribution of the location belonging to each category. The above distance calculation, similarity conversion and probability normalization process is performed on all spatial locations of the feature map to obtain a preliminary category probability map based on the prototype network. The spatial size of this probability map is the same as that of the feature map, containing 13 channels corresponding to the probabilities of 13 land cover categories.

[0039] Step 2.4: Fine-tune and optimize the encoder online using the support set; Based on the target domain support set and the encoder, an online fine-tuning strategy is employed to lightweight adjust the encoder parameters, making them better suited to the feature distribution of the current study region. The fine-tuning process uses the gradient descent optimization algorithm, and the classification loss function is defined as cross-entropy loss.

[0040] The loss function is constructed as follows: For each sample in the support set, features are extracted by the encoder and the predicted probability of its belonging to each category is calculated. Based on the true category of the sample, an indicator function is used to determine whether the true category is the current category. If it is, the value is 1; otherwise, the value is 0. The indicator function value is multiplied by the logarithm of the predicted probability of the category, and the sum is obtained over all categories to get the cross-entropy loss of the sample. Then, the average loss of all support samples is taken and the negative value is taken to obtain the total classification loss. This loss function measures the difference between the model's predicted probability and the true label. The smaller the loss value, the more accurate the prediction. To avoid overfitting, only the parameters of the last few layers of the encoder are fine-tuned, while the pre-trained parameters of the earlier layers are kept frozen. The learning rate is set to 0.001, and 10 to 20 rounds of iterative optimization are performed. In each iteration, the gradient of the loss function with respect to the trainable parameters is calculated, that is, the partial derivative of the loss function with respect to each parameter. The parameters are updated according to the gradient descent rule, specifically by subtracting the product of the learning rate and the gradient from the current parameter value to obtain the updated parameter value. This process adjusts the parameters along the direction of the fastest descent of the loss function. After multiple rounds of iterative optimization, an encoder adapted to the target domain is obtained. This encoder, while maintaining the pre-trained knowledge, is customized for the feature distribution of the current region.

[0041] Step 2.5: Extract target domain adaptation features using the adapted encoder; Based on the adapted encoder and the optimized image dataset, forward inference is used to re-extract deep semantic features from the images. The images are input into the adapted encoder, and the image data is propagated forward layer by layer through operations such as convolution, activation, and pooling. Feature extraction and transformation are performed sequentially in each layer of the encoder network. The output layer yields a feature map adapted to the target domain. The spatial dimensions of this feature map are the downsampled height and width, and the feature dimension is the number of feature channels.

[0042] Compared to the original embedded features, the adapted features better reflect the distribution of land cover features in the current study area because the encoder has been fine-tuned using the target domain support set and has learned specific feature patterns for the current area, exhibiting stronger discriminative ability for different land cover categories in the target domain. The adapted feature map serves as input for subsequent multi-scale feature extraction and fusion, providing an optimized feature representation for accurate identification of land cover in the current area.

[0043] In some embodiments, since the number of samples in the target domain support set is extremely small and the class distribution is unbalanced, a strategy combining data augmentation and prototype optimization can be adopted to improve model adaptability. Data augmentation operations such as random pruning, rotation, flipping, and color jitter are applied to the support set samples to generate multiple augmented versions for each original sample, expanding the support set size to 5 to 10 times its original size. The prototype vector is recalculated for the expanded support set using a weighted averaging strategy. Specifically, for a given class, all expanded samples of that class are traversed, and the feature vector of each sample is extracted by the encoder. The feature vector is multiplied by the corresponding sample weight, and the sum of all weighted feature vectors is calculated and then divided by the sum of all sample weights to obtain the weighted average prototype vector. The weight of the original sample is set to 1, and the weight of the augmented sample is set to 0.5 to 0.8 to reduce the influence of the augmented sample on the prototype and ensure that the prototype is mainly determined by the real samples. Simultaneously, an intra-class variance constraint is introduced to enhance the compactness of the prototype. A prototype compactness term is added to the loss function. This term is calculated as follows: for each class, all supporting samples of that class are traversed, the difference between the feature vector of each sample and the prototype vector of that class is calculated, the square of the L2 norm of the difference vector is calculated, the squared distances of all samples of that class are summed, and then the summation is performed for all classes to obtain the total compactness loss. This loss term encourages the clustering of features of similar samples towards the prototype, reducing the dispersion of intra-class features. By optimizing the joint objective function that includes classification loss and compactness loss, the stability and representativeness of the prototype representation are enhanced under the condition of very few samples, improving the model's rapid adaptability and classification accuracy.

[0044] Step 3: Based on the feature map adapted to the target domain, perform multi-scale dilated convolution and category-scale attention fusion to obtain a comprehensive fused feature map set; Step 3.1: Extract multi-scale features through multi-branch dilated convolution; Based on the feature map of the adaptive target domain, a multi-branch dilated convolutional network is used to extract feature representations at different receptive field scales. Dilated convolution expands the receptive field without increasing the number of parameters by inserting holes between the convolutional kernel elements. Four parallel dilated convolutional branches are set up with dilation rates of 1, 2, 4, and 8. The first branch has a dilation rate of 1, i.e., a standard 3×3 convolution, with a receptive field of 3×3, suitable for capturing small features and edge details; the second branch has a dilation rate of 2, with an effective receptive field of 5×5, suitable for capturing medium-scale features such as rural settlements; the third branch has a dilation rate of 4, with an effective receptive field of 9×9, suitable for capturing larger-scale features such as cultivated land; and the fourth branch has a dilation rate of 8, with an effective receptive field of 17×17, suitable for capturing large-scale features such as forest patches. Each branch contains two layers of dilated convolution, followed by batch normalization and ReLU activation. The adapted feature map is input into the four branches respectively, and four feature maps of different scales are obtained through parallel computation. The spatial dimensions of each feature map are the same as the input, and the number of channels is 256.

[0045] Step 3.2: Construct the correlation matrix between land cover categories and spatial scale; Based on a set of land cover categories and a pre-defined prior knowledge base of land cover scales, the typical spatial scale characteristics of each land cover category are obtained through a lookup table method. The prior knowledge base of land cover scales is established based on practical experience in land spatial planning and remote sensing interpretation standards, storing the typical patch size ranges for each land cover category. For example, transportation land use is mainly characterized by linear or narrow strip shapes, with a typical width of 5 to 30 meters, corresponding to small-scale characteristics; urban construction land use is characterized by medium-sized patches, with a typical size of 50 to 200 meters, corresponding to medium-scale characteristics; and forest land use is characterized by large, continuous patches, with a typical size of over 500 meters, corresponding to large-scale characteristics.

[0046] Based on the scale range in the knowledge base, 13 land cover categories are mapped to 4 dilated convolutional scales. For each category, the correlation degree between it and each scale is calculated. The closer the convolutional scale is to the typical scale of the category, the higher the correlation degree. The correlation degree is calculated based on the scale difference using a Gaussian function or a trigonometric function. A category-scale correlation matrix is ​​constructed. This matrix contains 13 rows corresponding to the 13 categories and 4 columns corresponding to the 4 dilated rates. The matrix elements represent the correlation strength between a certain category and a certain dilated rate, and the values ​​range from 0 to 1.

[0047] Step 3.3: Calculate the category-guided scale attention weights; Based on the category-scale correlation matrix and multi-scale feature map set, a category-guided scale attention mechanism is used to calculate the scale weight distribution that should be paid attention to for each category at each spatial location.

[0048] For each scale of feature map, a scale response map is generated. Specifically, global average pooling is performed on the feature map to average the feature values ​​of all spatial locations in each channel, resulting in a global feature vector. A 1×1 convolution is then applied to the global feature map to generate a two-dimensional scale response map, which reflects the activation intensity at each location of the feature map at that scale. For each category, its scale weight vector at each spatial location is calculated. This involves extracting the row vector corresponding to the category from the category-scale association matrix, which contains the association strength between the category and the four scales. The response values ​​of a specific spatial location on the four scale response maps are extracted and formed into a four-element vector. The category association vector and the location response vector are multiplied element-wise to obtain a score vector that integrates prior knowledge and data response. The softmax normalization function is applied to the score vector to calculate the exponent value of the score at each scale, and then divided by the sum of the exponent values ​​of all scales to obtain the normalized scale weight vector, ensuring that the sum of the weights at the four scales is 1. This calculation process integrates scale prior knowledge with data-driven attention: the category-scale correlation matrix provides prior guidance based on land cover characteristics, and the scale response map provides adaptive adjustment based on the current image data; the above scale weight calculation is performed on 13 categories respectively to obtain category-specific scale attention weight tensors, which contain weight information for 13 categories, 4 scales, and all spatial locations.

[0049] Step 3.4: Perform category-adaptive multi-scale feature fusion; Based on category-specific scale attention weights and multi-scale feature map sets, a weighted fusion strategy is used to construct optimized feature representations for each category.

[0050] For each category, its fused feature map is calculated. Specifically, at each spatial location, four scale weight values ​​corresponding to that category are extracted, and feature vectors of the four scale feature maps at that location are extracted respectively. The feature vector of each scale is multiplied by the corresponding weight value, and the four weighted feature vectors are added together to obtain the fused feature vector at that location. This fusion process customizes the optimal scale combination for each category. For small target categories such as transportation land, small-scale features receive higher weights, and for large patch categories such as woodland, large-scale features receive higher weights. The above fusion operation is performed on each of the 13 categories to obtain 13 category-specific fused feature maps. Based on the fused feature maps of all categories, a unified comprehensive fused feature representation is constructed using channel concatenation and convolutional dimensionality reduction operations. The fused feature maps of the 13 categories are concatenated along the channel dimension, that is, the 13 feature vectors are arranged sequentially at each spatial location to form a high-dimensional feature vector. The number of channels in the concatenated feature map is 256 multiplied by 13, which equals 3328. Dimensionality reduction is performed through 1×1 convolution. The convolution kernel performs linear combination and nonlinear transformation on all channels of the concatenated feature, compressing the 3328-dimensional feature to 512 dimensions, reducing the computational and storage overhead while retaining the optimized feature information of each category. The resulting comprehensive fused feature map set integrates multi-scale information and category prior knowledge, providing optimized feature representations for subsequent multi-source data fusion and classification.

[0051] In some embodiments, since the spectral characteristics of different land cover categories differ in different seasons, a phenologically adaptive scale attention adjustment strategy can be adopted. The seasonal category is determined based on the image imaging time, mapping the time to one of the four seasons: spring, summer, autumn, or winter. A three-dimensional correlation tensor of phenological-category-scale is established, containing the correlation relationships of the four seasons, 13 land cover categories, and 4 scales. This tensor stores the correlation strength between each land cover category and scale in different seasons. For example, the texture scale characteristics of deciduous forests change after losing leaves in winter, requiring adjustment of scale weight allocation. Different scale correlation values ​​are set for the winter forest category in the tensor. When calculating scale attention weights, a corresponding two-dimensional slice is selected from the three-dimensional tensor based on the seasonal index of the current image. This slice contains the correlation matrix of 13 categories and 4 scales for the current season. The seasonal-specific correlation matrix is ​​used instead of the general correlation matrix for subsequent scale weight calculations. This allows the scale attention mechanism to automatically adjust according to seasonal changes. For seasonally sensitive land cover categories such as deciduous forests and crops, different scale fusion strategies are adopted in different seasons. The purpose of this strategy is to enhance the method's adaptability to seasonal changes and improve the consistency and accuracy of image recognition results from different time periods.

[0052] Step 4: Evaluate the suitability score of the integrated feature map set and perform weighted fusion, classification, and normalization to obtain preliminary classification results; Step 4.1: Evaluate the adaptability of each data source to different land cover categories; Based on a comprehensive fusion feature map set and a set of land cover categories from different data sources, a data source-category fit evaluation network is used to quantify the ability of each data source to identify different land cover categories. The fit evaluation network adopts an attention mechanism structure, with the input being the fusion feature map and category prototype vector of a certain data source, and the output being the fit score of that data source for that category.

[0053] For a given data source and category, the prototype vector for that category is obtained from the category prototype set. The mean cosine similarity between all location features in the feature map of the data source and the prototype vector is calculated as the global fitness score. A small evaluation network is used to jointly encode the feature map and prototype vector. This network is trained on labeled samples from multiple data sources to learn the characteristics of different data sources. For example, high-resolution optical data has a strong ability to identify built-up land, Sentinel-2 data has a strong ability to classify vegetation, and SAR data has a strong ability to identify water bodies. The fitness score of the data source for that category is obtained through network inference. The score ranges from 0 to 1, with a higher score indicating that the data source is more suitable for identifying that category of land cover. Fitness scores are calculated for all combinations of all data sources and 13 categories, and a fitness matrix is ​​constructed. The number of rows in this matrix represents the number of data sources, the number of columns represents the 13 categories, and the matrix elements represent the fitness score of a data source for a specific category.

[0054] Step 4.2: Calculate the category-related multi-source fusion weight matrix; Based on the adaptability matrix and the multidimensional quality assessment vectors corresponding to each data source, a comprehensive weight calculation strategy is adopted to consider both data quality and category adaptability.

[0055] A base quality score is extracted from the quality assessment vector of each data source, which comprehensively reflects the physical quality of the data source. For each combination of data source and category, its fusion weight is calculated. Specifically, the base quality score of the data source is multiplied by its suitability score for that category to obtain the overall score of the data source for that category. The same overall score is calculated for all data sources, and the overall scores of all data sources are added together to obtain a normalized denominator. The overall score of the data source is divided by the normalized denominator to obtain the fusion weight of that data source for that category. This calculation logic ensures that high-quality and highly adaptable data sources receive a larger weight, while the weight of low-quality or poorly adaptable data sources is suppressed. The above weight calculation is performed on all combinations of data sources and categories to construct a category-related fusion weight matrix. The number of rows in this matrix is ​​the number of data sources, and the number of columns is 13 categories. Each column of the matrix corresponds to one category, and the sum of the column elements is 1, indicating the weight allocation of each data source when fusion within that category.

[0056] Step 4.3: Perform category-adaptive multi-source feature weighted fusion; Based on the fusion weight matrix and the fusion feature maps of each data source, a category-adaptive weighted fusion strategy is adopted to fuse multi-source features.

[0057] For each category, a weight vector corresponding to that category is extracted from the weight matrix. This vector contains the fusion weights of all data sources for that category. The multi-source fusion feature map corresponding to that category is then calculated. Specifically, all data sources are traversed, and the fusion feature map of each data source is multiplied by the weight of that data source for the current category to obtain a weighted feature map. The weighted feature maps of all data sources are then summed to obtain the multi-source fusion feature map for that category. This fusion operation is performed on each of the 13 categories to obtain 13 category-specific multi-source fusion feature maps. This fusion strategy implements an adaptive mechanism for selecting the optimal combination of data sources for different land cover categories. The fusion weights are automatically adjusted according to category characteristics. For example, SAR data receives higher weights when identifying water bodies, multispectral data receives higher weights when identifying vegetation, and high-resolution optical data receives higher weights when identifying buildings, thus fully leveraging the advantages of each data source in different land cover identification tasks.

[0058] Step 4.4: Generate a class probability map using a multi-head classifier; Based on the multi-source fusion feature maps of each category, a multi-head classifier architecture is adopted to independently construct a binary classifier for each category.

[0059] For each category, a dedicated binary classifier is constructed, consisting of a 1×1 convolutional layer and a sigmoid activation function. The convolutional layer performs a linear transformation on the 512-dimensional features, multiplying the feature vector by the weight matrix and adding a bias term to obtain a one-dimensional score. The sigmoid function performs a non-linear transformation on the score, mapping any real value to a probability value between 0 and 1, resulting in a probability map indicating that the location belongs to that category. This classification operation is performed on each of the 13 categories, resulting in 13 category probability maps. Based on the probability maps of all categories, a probability normalization operation is performed to obtain a category probability map that conforms to the probability distribution constraints. Specifically, at each spatial location, the probability values ​​of the 13 categories are extracted to form a probability vector. The sum of these 13 probability values ​​is calculated as the normalization denominator, and the probability value of each category is divided by this sum to obtain the normalized probability value, ensuring that the sum of the probabilities of all categories at that location is 1. The normalization operation is performed on all spatial locations to obtain the normalized category probability map. The probability map is maximized along the category dimension to obtain the preliminary classification result. That is, the category with the highest probability is selected as the classification label for each location, thus completing the pixel-level land cover classification.

[0060] In some embodiments, due to the spatial heterogeneity of the complementarity of multi-source data across different land cover categories, a locally adaptive fusion weight adjustment strategy can be adopted. The image is divided into multiple local region windows. Within each window, the distribution ratio of each category and the local quality indicators of each data source are statistically analyzed. The fusion weights are dynamically adjusted based on local statistical characteristics. A local adjustment function is used to calculate local adjustment coefficients. This function takes the local statistical characteristics of a location as input and outputs the adjustment coefficients of each data source for each category at that location. The global fusion weights are multiplied by the local adjustment coefficients to obtain the corrected local fusion weights. The weights of all data sources are normalized to ensure that the sum of the weights is 1. This strategy adaptively adjusts the data source weights based on local features within the image. For example, in areas partially obscured by clouds or fog, the weight of obscured data sources is reduced, while the weight of cloudless data sources is increased. In complex terrain areas, the weight of high-resolution data sources is increased to obtain more detailed information. In flat areas, more reliance can be placed on low-to-medium resolution data sources with rich spectral information. The purpose of this strategy is to achieve more refined data source selection, adapting to quality differences and land cover distribution characteristics within the image, further improving the fusion effect and recognition accuracy.

[0061] Step 5: Based on the preliminary classification results, change detection and segmentation of stable and changing regions are performed, temporal context features are extracted and constraint optimization is carried out to obtain the temporally optimized classification results; Step 5.1: Identify regions of significant change using a change detection network; Based on image data sequences from multiple time phases and corresponding preliminary classification result sequences, a twin change detection network is used to identify regions where significant changes occur between adjacent time phases.

[0062] For two adjacent temporal phases, the optimized images of the two phases are input into two branches of the Siamese network. Each branch uses the same convolutional encoder to extract features. The difference features between the two temporal feature maps are calculated by subtracting them element-wise to obtain the difference feature map. The absolute value of the difference is then taken to obtain the difference feature map. The difference feature map is input into the change detection decoder. The decoder restores the spatial resolution through upsampling and convolution operations and outputs a change probability map. The spatial size of this probability map is the height multiplied by the width of the image, and the value ranges from 0 to 1. The value of each pixel in the probability map represents the probability of a change in land cover at that location. The closer the value is to 1, the greater the probability of a change. The change detection network is trained on a dataset containing real land cover change annotations to learn the feature patterns of different types of land cover changes, such as farmland being converted into construction land or forest land being destroyed.

[0063] Step 5.2: Adaptively segment the changing region and the stable region; Based on the probability map of change, an adaptive threshold segmentation method is used to divide the image region into changing regions and stable regions.

[0064] Threshold selection employs either the Otsu's maximum inter-class variance method or a quantile-based method based on probability distribution, automatically determining the optimal segmentation threshold based on the histogram distribution of change probabilities. For each location, it is determined whether its change probability is greater than the segmentation threshold. If it is, the location is marked as a change region; otherwise, it is marked as a stable region. A binary mask for the change region is generated. An indicator function is used to determine whether the change probability is greater than the threshold; a 1 is output when the condition is true, and a 0 is output when the condition is false, indicating a stable region. The stable region mask is obtained by subtracting the change region mask from 1, making the change and stable regions complementary. Change detection and region segmentation are performed relative to all adjacent time intervals, resulting in a set of change masks for the time series. To avoid false detections and the influence of noise, morphological post-processing is performed on the change masks. Opening operations remove isolated noise points, and closing operations fill small holes in the change regions, resulting in smooth region masks.

[0065] Step 5.3: Extract the temporal context features of the stable region; Based on the multi-temporal preliminary classification result sequence, the corresponding category probability map sequence, and the stable region mask sequence, a bidirectional gated recurrent unit network is used to extract temporal context features.

[0066] For each time phase, the category probability map is multiplied element-wise with the stable region mask. That is, at each spatial location, the probability value is multiplied by the mask value. If the mask value is 0, the probability is set to zero; if the mask value is 1, the probability remains unchanged. This masks the probability values ​​of changing regions, retaining only the classification information of stable regions, thus obtaining the stable region probability map. The stable probability map of the time series is input into a bidirectional GRU network for time series modeling. The forward GRU processes data sequentially from the first time step to the last. At each time step, the forward GRU receives the stable probability map of the current time step and the forward hidden state of the previous time step as input, and updates the hidden state through a gating mechanism to capture the influence of historical information on the current time step. The backward GRU processes data in reverse order from the last time step to the first time step. At each time step, the backward GRU receives the stable probability map of the current time step and the backward hidden state of the next time step as input, and updates the hidden state through a gating mechanism to capture the constraints of future information on the current time step. At each time step, the forward hidden state and the backward hidden state are concatenated along the feature dimension to obtain a context feature that integrates bidirectional temporal information. The dimension of this feature is the sum of the dimensions of the forward hidden state and the backward hidden state. The temporal context feature encodes the position information of the time phase in the time series and its correlation with adjacent time phases, providing rich temporal semantic information for subsequent temporal consistency constraints.

[0067] Step 5.4, Regional adaptive temporal constraint optimization; Based on temporal context features, the class probability map of the current time phase, and the change region mask, a region-adaptive temporal constraint strategy is adopted to optimize the classification results.

[0068] The association strength between the current time phase and adjacent time phases is calculated using an attention mechanism. For the current time and its neighboring time phases (including the two previous time phases, the previous time phase, the next time phase, and the two next time phases), the process of calculating the attention weight involves performing a dot product operation between the context features of the current time and the context features of the neighboring time phases, then scaling the result by the square root of the hidden state dimension to obtain a similarity score. An exponential value is calculated for the scores of all neighboring time phases, and the exponential value of the current neighboring time is divided by the sum of the exponential values ​​of all neighboring time phases to obtain a normalized attention weight. This weight reflects the degree of influence of the classification results of the neighboring time phases on the current time phase, with higher similarity time phases receiving larger weights. Different constraint strengths are set according to the region type. For stable regions, a strong temporal consistency constraint is applied, with a constraint strength coefficient set to 0.5. For changing regions, a weak temporal constraint or no constraint is applied, with a constraint strength coefficient set to 0.1. A position-related constraint strength map is constructed. At each position, the strong constraint coefficient is multiplied by the mask value of the stable region, and then multiplied by the weak constraint coefficient by the mask value of the changing region to obtain the constraint strength at that position. The time-series optimized category probability map is calculated. For each position and each category, a weighted average adjustment of the probabilities of adjacent time phases is calculated. Specifically, the neighboring time phases are traversed, the difference between the probability of that category at each neighboring time phase and the probability at the current time phase is calculated, the difference is multiplied by the corresponding attention weight, the weighted differences over all neighboring time phases are summed, and then multiplied by the constraint strength coefficient at that position to obtain the adjustment amount. The original probability at the current time phase is added to the adjustment amount to obtain the optimized probability. This calculation process achieves weighted smoothing of the classification results of adjacent time phases, applies strong constraints to stable regions to maintain consistency in the time series, and applies weak constraints to changing regions to allow reasonable changes. The optimized probability map is then maximized in the category dimension, and the category with the highest probability is selected as the classification label for that position, resulting in the time-series optimized classification result.

[0069] In some embodiments, since certain real-world feature changes are sudden and have clear semantic transformation patterns, a strategy combining change semantic modeling and constraint adjustment can be adopted. A feature category transformation rule base is established, storing reasonable category transformation paths and their probabilities of occurrence. For example, converting farmland to construction land is a reasonable high-probability transformation, but directly converting water bodies to forest land is an unreasonable low-probability transformation. In the changed area, the category transformation types of adjacent time phase classification results are analyzed, and the category labels of the current and next time phases are extracted. The rule base is queried to determine the reasonableness of the transformation, obtaining a transformation reasonableness score. The higher the score, the more reasonable the transformation. The constraint strength is dynamically adjusted based on the transformation reasonableness score. For reasonable transformations, a weak constraint strength is maintained, allowing the change to occur; for unreasonable transformations, the constraint strength is increased or a manual review flag is triggered. The specific constraint strength adjustment calculation involves calculating the complement of the reasonableness score (i.e., 1 minus the reasonableness score), multiplying the complement by the difference between the strong constraint coefficient and the weak constraint coefficient, and adding the weak constraint coefficient to obtain the adjusted constraint strength. This calculation logic ensures that when the reasonableness score is high, the constraint strength is close to the weak constraint coefficient, and when the reasonableness score is low, the constraint strength is close to the strong constraint coefficient. The purpose of this strategy is to suppress unreasonable category jumps caused by misidentification while allowing changes in real-world features. It improves the logical consistency and credibility of the time series results by intelligently adjusting temporal constraints through semantic rules.

[0070] Step 6: Based on the classification results after time-series optimization, calculate the information entropy and the uncertainty of the decision boundary, and obtain the confidence score by weighted fusion according to the change region; Step 6.1: Calculate the uncertainty based on information entropy and decision boundary; Based on the time-series optimized classification results, the information entropy method is used to quickly quantify the uncertainty of classification. For each spatial location, its category probability vector is extracted, which contains the probability values ​​of that location belonging to 13 categories. The information entropy of that location is calculated by multiplying the probability value of that category by its logarithm for each category, summing the products of all categories, and taking the negative value. The entropy value reflects the uniformity of the probability distribution. When the probability of one category is close to 1 and the probabilities of other categories are close to 0, the entropy value is small, indicating high classification certainty. When the probabilities of multiple categories are similar, the entropy value is large, indicating greater classification uncertainty. The entropy value is calculated for all locations to obtain an uncertainty map based on entropy. Simultaneously, the maximum probability difference method is used to evaluate the distance to the decision boundary. For each location, the highest probability and the second highest probability are extracted from the probability vector. Specifically, all categories are traversed to find the category with the highest probability and its probability value as the highest probability, and then the category with the second highest probability and its probability value are found as the second highest probability. The difference between the highest probability and the second highest probability is calculated. The larger the difference, the greater the probability gap between the optimal category and the second best category, and the clearer the classification decision. The smaller the difference, the more multiple candidate categories are competing, and the higher the classification uncertainty. The difference is converted into an uncertainty measure. The uncertainty value is obtained by subtracting the difference from 1, resulting in an uncertainty map based on the decision boundary.

[0071] Step 6.2: Estimate model uncertainty through lightweight deep ensemble; Based on a comprehensive fusion of feature maps, a lightweight deep ensemble approach is used to approximate the uncertainty in model predictions. Traditional deep ensemble methods require training multiple independent complete models, resulting in enormous computational overhead. This embodiment employs a lightweight architecture with a shared feature extractor and a five-classifier head, diversifying only the classification layer.

[0072] Specifically, the parameters of the feature extraction encoder and the feature fusion network are kept fixed. Five independent classification head networks are constructed in the classification stage. Each classification head contains two fully connected layers and a softmax output layer. Through different random initializations and different training data sampling, each classification head learns a variety of classification boundaries.

[0073] During the inference phase, the fused feature map is input into five classifiers, each performing independent forward inference to obtain five class probability predictions. For each position and each class, the probability predictions from the five classifiers are collected, and the average of these predictions is calculated as the average probability (by summing the five probability values ​​and dividing by 5). The variance of the probability predictions is calculated by taking the difference between each classifier's prediction and the average probability, squaring the difference, summing the squared differences from all classifiers, and dividing by 5 to obtain the prediction variance for that class at that position. The variances of all 13 classes are averaged, summed, and divided by 13 to obtain the model uncertainty at that position. Variance reflects the dispersion of predictions from different classifiers; a large variance indicates internal discrepancies in the model's classification at that position, indicating high uncertainty, while a small variance indicates consistent predictions from all classifiers, indicating low uncertainty.

[0074] Step 6.3: Integrate multidimensional uncertainties and consider adjustments for changing regions; Based on entropy uncertainty, decision boundary uncertainty, model uncertainty, and change region mask, a multi-dimensional uncertainty fusion strategy is adopted to calculate the comprehensive uncertainty score.

[0075] Three uncertainty measures are normalized by mapping each measure to the interval between 0 and 1 using max-min normalization. Specifically, for each uncertainty graph, its minimum and maximum values ​​are found. The uncertainty value at each position is subtracted from the minimum value and then divided by the difference between the maximum and minimum values ​​to obtain the normalized uncertainty value. The three normalized uncertainties are then fused using a weighted summation method. The entropy uncertainty is multiplied by a weight coefficient of 0.4, the decision boundary uncertainty by a weight coefficient of 0.3, and the model uncertainty by a weight coefficient of 0.3. The three weighted values ​​are then added together to obtain the fused uncertainty. The weight coefficients are determined based on experience or cross-validation and reflect the relative importance of each uncertainty measure. Considering the typically high uncertainty in identifying change regions, a region-related uncertainty adjustment strategy is adopted. An uncertainty weight map is constructed. At each location, the uncertainty enhancement coefficient of the change region (0.5) is multiplied by the mask value of the change region, and then 1 is added to obtain the uncertainty weight for that location. For stable regions with a mask value of 0, the weight is 1; for change regions with a mask value of 1, the weight is 1.5. The fusion uncertainty is weighted by multiplying the fusion uncertainty value at each location by its corresponding weight to obtain the final uncertainty, thus relatively increasing the uncertainty score of the change region and giving it higher priority in subsequent manual verification. The comprehensive uncertainty is converted into a confidence score. The minimum and maximum values ​​of the final uncertainty map are found. For each location, the minimum value is subtracted from its uncertainty value, and then divided by the difference between the maximum and minimum values ​​to obtain a normalized uncertainty value. This normalized value is subtracted from 1 to obtain the confidence score, which ranges from 0 to 1. A higher confidence score indicates a more reliable identification result at that location.

[0076] In some embodiments, due to the varying identification difficulty of different land cover categories, a category-related uncertainty calibration strategy can be adopted. Based on the validation dataset, the confusion matrix and identification accuracy of each category are statistically analyzed to establish a category difficulty table. This table stores the difficulty coefficients for each of the 13 land cover categories. The difficulty coefficients are positively correlated with the category error rate; categories with higher error rates have larger difficulty coefficients. When calculating the overall uncertainty, a category difficulty adjustment is introduced. For each location, the classification result (predicted category label) is obtained, and the corresponding difficulty coefficient for that category is retrieved from the category difficulty table. A difficulty adjustment factor is calculated by multiplying the difficulty adjustment weight (0.3) by the difficulty coefficient and adding 1. The final uncertainty value for that location is then multiplied by the adjustment factor to obtain the calibrated uncertainty. This strategy increases the uncertainty score for difficult-to-identify categories such as shrublands and orchards, even if the model gives a high probability, prompting manual verification of these easily confused categories. For easily identifiable categories such as water bodies and construction land, the uncertainty score remains low. The purpose of this strategy is to integrate prior knowledge of categories into uncertainty assessment, improve the rationality of prioritizing verification tasks, and prioritize limited manual verification resources for high-risk areas.

[0077] Step 7: Based on the confidence score, perform conditional random field optimization, confidence leveling and connected component identification to generate a priority-ranked verification task list; Step 7.1: Optimize spatial consistency using conditional random fields; Based on the time-optimized classification results, category probability maps, and available region masks, a Conditional Random Field (CRF) model is used to perform spatial consistency post-processing on the classification results within the available region. The CRF model models the classification problem as an energy minimization problem, with the energy function consisting of two parts: a univariate potential function and a bivariate potential function.

[0078] An energy function is constructed, consisting of two parts. The first part is a univariate potential function, which is summed over all pixels. The univariate potential function for each pixel measures the cost of assigning the pixel to a certain category. Specifically, it is calculated by extracting the probability that the pixel belongs to that category, calculating the logarithm of the probability and taking the negative value to obtain the cost. The higher the probability, the lower the cost. The second part is a binary potential function, which is summed over all adjacent pixel pairs. The binary potential function for each pixel pair measures the penalty for adjacent pixels not belonging to the same category. The binary potential function employs the Potts model. The specific calculation process involves determining whether two adjacent pixels belong to different categories; if they do, the indicator function is set to 1, otherwise it is set to 0. The difference between the image feature vectors of the two pixels is calculated, and the square of the L2 norm is calculated on the difference vector. This squared value is divided by twice the square of the feature similarity parameter, and the negative value is used to calculate the exponential function, yielding the feature similarity weight. The indicator function value, the feature similarity weight, and the smoothing intensity parameter are multiplied to obtain the binary potential function value. This design imposes a larger penalty when adjacent pixels belong to different categories but have similar image features, encouraging adjacent pixels with similar features to be assigned to the same category, thus achieving spatial smoothing. The smoothing intensity parameter controls the strength of the spatial constraint, while the feature similarity parameter controls the sensitivity of feature similarity. An efficient approximate inference algorithm, such as the mean field approximation or graph cut algorithm, is used to minimize the energy function, iteratively updating the pixel categories until convergence, resulting in a spatially optimized classification result. Conditional random field optimization can eliminate isolated noise points, smooth ground object boundaries, and improve the visual quality and spatial coherence of the classification results.

[0079] Step 7.2, confidence level classification and identification of areas to be verified; Based on the recognition confidence map, a multi-threshold grading strategy is adopted to divide the recognition results into different levels according to their confidence.

[0080] Set a high confidence threshold of 0.8 and a low confidence threshold of 0.6; generate three levels of masks; the high confidence region mask uses an indicator function to determine whether the confidence level is greater than or equal to the high confidence threshold, outputting 1 if the condition is true and 0 otherwise; the recognition results for this region are highly reliable and can be used directly; the medium confidence region mask uses an indicator function to determine whether the confidence level is greater than or equal to the low confidence threshold and less than the high confidence threshold, outputting 1 if the condition is true and 0 otherwise; this region needs to be verified in conjunction with other auxiliary information; the low confidence region mask uses an indicator function to determine whether the confidence level is less than the low confidence threshold, outputting 1 if the condition is true and 0 otherwise; this region has greater uncertainty and requires manual verification.

[0081] Based on a low-confidence region mask, connected component analysis is used to identify specific land parcels requiring manual verification. A connected component labeling algorithm segments low-confidence regions into multiple independent connected regions. For each connected region, attributes such as area, perimeter, average confidence score, and relevant category are calculated. Connected regions with an area greater than a preset minimum threshold (e.g., 100 square meters) are selected as valid verification units, filtering out excessively small isolated noise points. The selected verification units are sorted from low to high average confidence score, with lower-confidence units having higher priority. A list of regions to be verified is generated, containing detailed attributes for each verification unit, including spatial location, area, current classification category, average confidence score, and uncertainty score, providing clear guidance for verification tasks to business personnel.

[0082] Step 7.3, Vectorization Conversion and Generation of Result Data; Based on the spatially optimized classification results and confidence level mask, a raster-to-vector algorithm is used to convert the classification results into a vector format that conforms to the national land spatial planning data standard.

[0083] The vectorization process extracts the boundary contours of each land cover category and identifies category boundaries using edge detection and contour tracking algorithms. The extracted contours are simplified using the Douglas-Puk algorithm to reduce redundant vertices while preserving key morphological features. The simplified contours are then converted into vector polygon objects, with each polygon corresponding to a land cover patch. Attribute information is added to the vector polygons, including land cover category code, patch area, perimeter, average confidence level, and quality grade. The vector data structure is organized according to national land spatial planning data standards, and vector output data is generated in Shapefile or GeoPackage format.

[0084] Based on vector data and confidence level information, thematic map products are generated using automated mapping tools. Thematic maps use different colors to represent different land cover categories, and fill styles or transparency indicate confidence levels: high-confidence areas are filled with solid fills, medium-confidence areas with semi-transparent fills, and low-confidence areas with grid fills or bolded boundary markings. The spatial distribution of areas to be verified is overlaid on the thematic map and highlighted with red boundaries or special symbols. A quality assessment report is generated, summarizing the identified area, confidence level distribution, and data source usage for each land cover category, providing decision-making references for planning management departments. Prioritized verification task lists are generated, arranging areas to be verified from highest to lowest uncertainty, and marking the coordinates, image time, and suspected problem type for each area, supporting staff in efficiently conducting on-site verification work.

[0085] In some embodiments, since land spatial planning operations need to interface with planning databases, an incremental update and change feature generation strategy can be adopted. The identification results of the current time phase are spatially overlaid with historical data from the previous time phase to identify newly added, reduced, and category-transformed land feature features. For newly added features, their spatial extent and attribute information are extracted and marked as objects to be added to the database; for reduced features, they are marked as objects to be verified for deletion; for category-transformed features, information such as the category before and after the transformation, the transformation time, and the transformation area are extracted to generate thematic data on change features. Change feature data is organized according to the requirements for monitoring illegal and irregular land use in land spatial planning, supporting the rapid identification of issues such as the "non-agriculturalization" of cultivated land, encroachment on ecological protection red lines, and illegal construction. Change feature data is associated with confidence level information, prioritizing the push of high-confidence change features into the enforcement process, while low-confidence change features are included in verification task sheets for manual confirmation. The aim is to achieve seamless integration of identification results with the business system, improving the automation level and response efficiency of land spatial planning monitoring.

[0086] An adaptive recognition system for remote sensing images in land spatial planning is provided to perform the aforementioned adaptive recognition method for remote sensing images in land spatial planning. Figure 2 As shown, it includes: The data acquisition and preprocessing module is used to acquire multi-source remote sensing image datasets of the study area, perform feature extraction, cloud detection and quality assessment, and adaptive preprocessing to obtain optimized image datasets. The prototype network adaptive module, based on the optimized image dataset, performs prototype network encoding, calculates category prototypes and probabilities, and fine-tunes the support set to obtain feature maps adapted to the target domain. The multi-scale feature fusion module performs multi-scale dilated convolution and category-scale attention fusion based on the feature map adapted to the target domain to obtain a comprehensive fusion feature map set; The multi-source data fusion and classification module is used to evaluate the suitability score of the integrated fusion feature map set and perform weighted fusion, classification and normalization to obtain preliminary classification results; The temporal consistency optimization module, based on the preliminary classification results, detects changes and segments stable and changing regions, extracts temporal context features and performs constraint optimization to obtain the temporally optimized classification results. The uncertainty assessment module calculates the information entropy and uncertainty of the decision boundary based on the time-series optimized classification results, and obtains the confidence score by weighted fusion according to the change region. The post-processing and result generation module performs conditional random field optimization, confidence leveling, and connected component identification based on confidence scores, and generates priority-ranked verification task sheets.

[0087] In one embodiment of the present invention, a specific example is provided: A 365-day field test was conducted in a territorial spatial planning monitoring area of ​​a county in the southwestern mountainous region. During the test, a multi-source remote sensing data acquisition system was deployed, including data from the Gaofen-2 satellite, Sentinel-2 satellite, Zhuhai-1 commercial satellite, and an unmanned aerial vehicle (UAV) aerial survey platform. The test area covered 2,800 square kilometers, encompassing complex mountainous terrain ranging from 800 to 4,200 meters in altitude. Data acquisition spanned all four quarters of 2024, acquiring multi-source remote sensing imagery data across 12 temporal phases. Simultaneously, 130 ground-based measured sample points were collected for model training and accuracy verification. During the test, a systematic evaluation was conducted on the spatial resolution, spectral characteristics, temporal distribution, and quality of different data sources, and a comprehensive statistical analysis was performed on the identification effectiveness of 13 types of territorial spatial planning features.

[0088] Examples of multi-source remote sensing image data acquisition are shown in Table 1: Table 1: Examples of multi-source remote sensing image data acquisition; Table 2 shows an example of how to obtain target feature category samples: Table 2: Examples of target feature category sample acquisition; Through the above implementation methods, the present invention realizes intelligent adaptive recognition of remote sensing images for land and space planning, and solves key technical problems such as image quality differences, scarcity of training samples, heterogeneity of multi-source data, and temporal consistency, providing an efficient and reliable technical means for the current status investigation, planning preparation, and implementation supervision of land and space planning.

[0089] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for adaptive identification of remote sensing images for territorial space planning, characterized in that, The method comprises the following steps: obtaining a multi-source remote sensing image dataset of a research area, performing feature extraction, cloud detection and quality evaluation, and self-adaptive preprocessing to obtain an optimized image dataset; based on the optimized image dataset, performing prototype network coding, calculating category prototypes and probabilities, and supporting set fine-tuning to obtain a feature map adapted to the target domain; based on the feature map adapted to the target domain, performing multi-scale hollow convolution and category scale attention fusion to obtain a comprehensive fusion feature map set; evaluating the adaptability score of the comprehensive fusion feature map set and performing weighted fusion, classification and normalization to obtain a preliminary classification result; based on the preliminary classification result, performing change detection and segmenting stable and change areas, extracting temporal context features and performing constraint optimization to obtain a temporally optimized classification result; based on the temporally optimized classification result, calculating the information entropy and the uncertainty of the decision boundary, and according to the change area, performing weighted fusion to obtain a confidence score; based on the confidence score, performing conditional random field optimization, confidence grading and connected component recognition to generate a priority-ordered verification task list. 2.The method of claim 1, wherein, The method for obtaining the optimized image dataset specifically comprises: using a multi-scale convolutional neural network to extract features in parallel on three feature channels of spectrum, texture and spatial structure to obtain multi-dimensional feature representation; based on the multi-dimensional feature representation and a digital elevation model, using a cloud detection special deep learning model to identify cloud-shielded areas and calculate cloud coverage; using a no-reference image quality evaluation network to extract a basic quality vector, and using a task relevance evaluation module to calculate the task quality score of each feature category; based on the basic quality vector, the task quality score and the cloud coverage, constructing a multi-dimensional quality evaluation vector, and determining a preprocessing parameter vector through an adaptive preprocessing strategy selection network; performing radiation correction, contrast enhancement, denoising, sharpening and cloud and fog removal processing according to the preprocessing parameter vector to obtain the optimized image dataset. 3.The method of claim 1, wherein, The method for obtaining the feature map adapted to the target domain specifically comprises: performing deep feature coding on the optimized image dataset through a pre-trained prototype network encoder to obtain an embedded feature map; based on a target domain support set, calculating the mean vector of the support sample feature vector of each feature category as a category prototype vector to form a prototype set; calculating the Euclidean distance between the feature vector of each spatial position in the embedded feature map and each category prototype vector, converting the distance into a similarity score and performing normalization to obtain a category probability map; based on the target domain support set, using a gradient descent optimization algorithm to fine-tune the encoder parameters and adjust the last preset number of layers of the encoder; using the adapted encoder to re-extract image features to obtain a feature map adapted to the target domain.

4. The method of claim 1, wherein the method is a method of self-adaptive identification of remote sensing images for territorial space planning. The method for obtaining the comprehensive fusion feature map set specifically comprises: using a multi-branch hollow convolution network to extract feature representation of different receptive field scales, and setting four parallel hollow convolution branches; based on a feature scale prior knowledge base, constructing an association matrix of feature categories and spatial scales, and the association matrix representing the association strength of each feature category and each hollow rate; generating a scale response map for each scale feature map, combining the category scale association matrix and the scale response map, and calculating a scale weight vector of each category at each spatial position; For each category, the feature maps of the four scales are fused by weighting according to the corresponding scale weight to obtain a category-specific fused feature map; The fused feature maps of all categories are channel-spliced and convoluted to obtain a set of comprehensive fused feature maps.

5. The method of claim 1, wherein the method is a method of self-adaptive identification of remote sensing images for territorial space planning. The obtaining of the preliminary classification result specifically includes: The input is the fused feature map of the data source and the category prototype vector, and the data source category adaptability evaluation network is used to quantify the recognition ability of each data source for different ground object categories, and an adaptability matrix is constructed; Based on the adaptability matrix and the multi-dimensional quality evaluation vector of each data source, the basic quality score of the data source is multiplied by its adaptability score for each category and normalized to construct a category-related fusion weight matrix; For each category, the fused feature maps of each data source are fused by weighting according to the corresponding fusion weight to obtain a category-specific multi-source fused feature map; A dedicated binary classifier is constructed for each category, and the multi-source fused feature map is classified to obtain a category probability map. The probability maps of all categories are normalized and the maximum value is taken to obtain a preliminary classification result.

6. The method of claim 1, wherein the method is a method of self-adaptive identification of remote sensing images for territorial space planning. The obtaining of the time-optimized classification result specifically includes: A twin change detection network is used to identify the significant change area between adjacent time phases, calculate the difference features of the two time phase feature maps, and output a change probability map; An adaptive threshold segmentation method is used to divide the image area into a change area and a stable area to generate a change area mask and a stable area mask; The category probability map of each time phase is multiplied by the stable area mask, and a bidirectional gated recurrent unit network is used to extract the time sequence context features; The correlation strength between the current time phase and the adjacent time phase is calculated through an attention mechanism, different constraint strength coefficients are set according to the region type, a strong time sequence consistency constraint is applied to the stable area, and a weak time sequence constraint is applied to the change area, to obtain a time-optimized classification result.

7. The method of claim 1, wherein the method is a method of self-adaptive identification of remote sensing images for territorial space planning. The obtaining of the confidence score specifically includes: The information entropy of the category probability vector of each spatial position is calculated to obtain an entropy-based uncertainty map; The difference between the highest probability and the second highest probability is calculated and converted into an uncertainty measure to obtain a decision boundary-based uncertainty map; A lightweight deep integrated architecture with a shared feature extractor and a five-classification head is used to calculate the variance of the prediction results of the five classification heads to obtain the model uncertainty; The three uncertainty measures are normalized and weighted summed, the fusion uncertainty is region-related adjusted according to the change area mask, and the adjusted uncertainty is converted into a confidence score. 8.The method of claim 1, wherein the method is a method of remotely sensing image adaptive identification for territorial space planning. The generation of the priority-ordered verification task list specifically includes: A conditional random field model is used for spatial consistency post-processing of the classification result to construct an energy function containing a unary potential function and a binary potential function, and the spatial-optimized classification result is obtained by minimizing the energy function; High confidence threshold and low confidence threshold are set to divide the recognition result into high confidence area, medium confidence area and low confidence area; Connected component analysis is performed on the low confidence area, and connected components with an area greater than a preset minimum threshold are selected as verification units and sorted from low to high according to the average confidence; A raster-to-vector algorithm is used to convert the classification result into a vector format to generate a priority-ordered verification task list. 9.The method of claim 2, wherein the method further comprises: The adaptive preprocessing strategy selection network includes: ​ The denoising parameters are automatically adjusted based on the signal-to-noise ratio score, and strong denoising parameters are selected for images with low signal-to-noise ratio. Adjust the contrast enhancement parameters according to the contrast score, and select strong contrast enhancement parameters for images with insufficient contrast. Whether to enable the dehazing algorithm is determined based on the cloud coverage level; for images with high cloud coverage, dehazing processing is automatically enabled. For high-quality images, use lightweight processing to avoid over-enhancement, and for low-quality images, use enhanced processing to improve data usability. 10.A system for adaptive identification of remote sensing images for territorial space planning, characterized in that, An adaptive remote sensing image recognition method for performing any one of claims 1-9 of a land spatial planning method includes: The data acquisition and preprocessing module is used to acquire multi-source remote sensing image datasets of the study area, perform feature extraction, cloud detection and quality assessment, and adaptive preprocessing to obtain optimized image datasets. The prototype network adaptive module, based on the optimized image dataset, performs prototype network encoding, calculates category prototypes and probabilities, and fine-tunes the support set to obtain feature maps adapted to the target domain. The multi-scale feature fusion module performs multi-scale dilated convolution and category-scale attention fusion based on the feature map adapted to the target domain to obtain a comprehensive fusion feature map set; The multi-source data fusion and classification module is used to evaluate the suitability score of the integrated fusion feature map set and perform weighted fusion, classification and normalization to obtain preliminary classification results; The temporal consistency optimization module, based on the preliminary classification results, detects changes and segments stable and changing regions, extracts temporal context features and performs constraint optimization to obtain the temporally optimized classification results. The uncertainty assessment module calculates the information entropy and uncertainty of the decision boundary based on the time-series optimized classification results, and obtains the confidence score by weighted fusion according to the change region. The post-processing and result generation module performs conditional random field optimization, confidence leveling, and connected component identification based on confidence scores, and generates priority-ranked verification task sheets.

Images