Ground feature classification method and system based on observation window and input feature optimization

Abstract

The application discloses a kind of based on observation window and input feature optimization ground object classification method and system, method includes S1 reading time phase observation window and input feature subset, and obtain the time phase remote sensing optical and SAR image in the time phase observation window of complete growth season corresponding time phase of study area;S2 according to input feature subset, extract the feature of remote sensing optical and SAR image, and input network model to obtain the time phase attention weight of each time phase;S3 calculate the confidence of time phase observation window;S4 it is judged whether the confidence of time phase observation window is greater than preset threshold, if yes, then enter S6, otherwise enter step S5;S5 at least one time phase is shifted to left and right with time phase observation window, obtain two alternative windows, calculate the window score of two alternative windows, using the alternative window of maximum window score updates time phase observation window, and enter S6;S6 optical and SAR feature corresponding to all time phase observation windows are input into XGBoost model, and the fine class map of multiple ground objects is obtained.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and land cover classification, specifically involving a land cover classification method and system based on the optimization of observation window and input features. Background Technology

[0002] Large-scale, high-precision land cover and crop type mapping are fundamental to food security assessment and agricultural resource management. In practical operational scenarios, traditional methods relying on dense temporal imagery face severe "data cost walls" and "computational efficiency walls" when dealing with complex classification tasks at the provincial and even national scales.

[0003] Current methods suffer from the following main shortcomings: First, high data costs. Downloading and processing dozens or even hundreds of images results in extremely high storage and computational overhead. Second, feature redundancy and overfitting, along with an imbalance in accuracy between categories. Training with all features introduces significant noise; as the number of categories increases and sample distribution becomes uneven, the discriminative features of weaker categories are overwhelmed, leading to a severe imbalance in classification accuracy. Third, the critical window relies heavily on experience and is sensitive to missing data. Determining the most important temporal window for classification heavily depends on expert prior knowledge, making automated implementation difficult in regions with complex climates and cropping systems. Furthermore, the critical identification period for crops often overlaps with the rainy season, resulting in frequent missing optical images; methods relying solely on optical data are almost ineffective under these conditions.

[0004] In recent years, attention mechanisms have been introduced into remote sensing time-series classification, but they are mostly used as "black box" modules to improve accuracy, with very few explicitly transforming weights into operable observation windows. Synthetic Aperture Radar (SAR) data has all-weather observation capabilities, but existing fusion of optical and SAR data mostly stays at the level of feature stitching or hidden layer fusion, rarely incorporating the physical response relationship between optical vegetation index and SAR backscattering as an explicit constraint into model training, thus failing to deeply explore the temporal complementary relationship between the two. On the other hand, ensemble learning models, represented by XGBoost, can output feature importance ranking, but when faced with massive candidate features for multi-class classification, the selected features are often isolated on the time axis and biased towards the dominant category with large sample size, paying insufficient attention to the key discriminative features of the sparsely sampled category, resulting in low land cover classification accuracy.

[0005] Therefore, there is an urgent need for a lightweight method that can automatically establish cross-modal collaboration between SAR and optical systems, lock multiple flexible observation windows, and fairly screen features within the windows, so as to achieve all-weather, high-precision land cover classification at extremely low cost. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, the land cover classification method and system based on observation window and input feature optimization provided by this invention solves the problems of high data cost and low land cover classification accuracy in existing land cover classification methods.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: Firstly, a land cover classification method based on observation window and input feature optimization is provided, which includes the following steps: S1. Read the stored temporal observation window and input feature subset, and obtain remote sensing optical and SAR images within the temporal observation window corresponding to the complete growing season of the study area; S2. Based on the features in the input feature subset, extract the optical and SAR features of remote sensing optical and SAR images, and input them into the trained network model to obtain the temporal attention weights for each time phase. S3. Calculate the confidence level of each phase observation window based on all phase attention weights corresponding to each phase observation window. S4. Determine whether the confidence level of the phase observation window is greater than the preset threshold. If yes, proceed to step S6; otherwise, proceed to step S5. S5. Shift the phase observation window to the left or right by at least one phase to obtain two alternative windows. Calculate the window scores of the two alternative windows. Update the phase observation window with the alternative window with the largest window score and proceed to step S6. S6. Input the optical and SAR features corresponding to all temporal observation windows into the XGBoost model to obtain detailed category maps of various land features.

[0008] In one possible implementation, the methods for constructing the stored temporal observation window and the input feature subset include: A1. Acquire multi-temporal remote sensing optical and SAR images of ground features in the target area over several years, covering the complete growing season. Extract the optical and SAR features of each ground feature sample and combine them with the ground feature labels of the ground features to form a dual-modal time series sample set. A2. Divide the dual-modal temporal sample set into a training set and a test set, and use the training set to train the dual-stream cross-modal attention network to obtain a network model that extracts the temporal attention weights for each phase. A3. Input the samples in the test set into the network model to obtain the temporal attention weights for each sample at each time phase, and determine several temporal observation windows based on the temporal attention weights and cloud probabilities; A4. For samples in the dual-modal time series sample set, retain the temporal phase and all optical and SAR features that are within a defined observation window, and perform spectral clustering to remove duplicates from the retained features to obtain multiple candidate features. A5. Input all candidate features from the dual-modal time series sample set into the XGBoost model and output the cumulative gain of each candidate feature for the land cover category; A6. Based on the cumulative gain, each time the candidate feature with the largest comprehensive score for the subset is selected and added to the subset, and when the subset meets the termination condition, the subset is used as the input feature subset. A7. Determine whether at least one feature is selected into the feature subset within the time phase observation window or whether the time phase observation window has not changed in the last two times. If so, retain the most recent time phase observation window and the feature subset; otherwise, proceed to step A8. A8. Reduce the temporal attention weight of the window without features by a preset ratio, and return to step A4 after redetermining the temporal observation window.

[0009] In one possible implementation, methods for determining the temporal observation window based on temporal attention weights and cloud probabilities include: A31. Average the phase attention weights for the same phase to obtain the average phase attention vector for all phases; average the cloud probabilities for the same phase for all samples in the test set to obtain the average cloud probabilities for all phases. A32. Obtain the set of time scales. For each time scale, slide across the full-phase sequence of the complete growing season with a preset step size to generate all possible continuous phase windows. A33. Calculate the window score for each consecutive time phase window based on all average time attention vectors and average cloud probabilities within the consecutive time phase window. A34. Select the average temporal attention weights that meet the preset conditions from the average temporal attention vectors, mark their corresponding temporal phases as core temporal phases, and generate the neighborhood of each core temporal phase; A35. Filter out the phase windows located in the neighborhood among all continuous phase windows, and select the phase observation window with the highest window score in each neighborhood as the phase observation window.

[0010] The beneficial effects of the above technical solution are as follows: This solution improves the temporal attention weighting. Heyun Probability The average attention is obtained by averaging the attention over time. and average cloud probability This makes both time-phase statistics, allowing them to be directly multiplied and weighted in subsequent window scoring. The multi-scale sliding window uses a set of time scales. Candidate windows, combined with quality weights Each window is scored to optimize the selection process, balancing information density and data usability. Core temporal phases are detected using a dual-condition approach: "local maxima + neighborhood saliency." Filter windows from the rating database within the neighborhood and apply a length penalty. Selecting the optimal observation window. This series of operations allows the extracted observation window to tolerate interannual shifts in the reproductive period and adapt to changes in window width, solving the problem of decreased accuracy of traditional fixed threshold windows in cross-year and cross-regional applications.

[0011] In one possible implementation, the preset condition is that the average temporal attention simultaneously satisfies both condition A and condition B: Condition A is and , , and Phases , and Average temporal attention; condition B is Higher than the neighborhood The mean of the average temporal attention plus 1.5 times the standard deviation of the neighborhood; The expression for selecting the time series window with the highest window score as the time phase observation window within the same neighborhood is:

[0012] in, For the i-th neighborhood The selected timing window; For the i-th core phase The neighborhood; Time scale The next A continuous time window Window rating; This is the length adjustment factor; This is a penalty item based on the scale.

[0013] The beneficial effects of the above technical solution are: Condition A ensures the marked core time phase. This is a local maximum point on the attention curve. Condition B further requires that this peak value be statistically significant within its neighborhood, thereby effectively suppressing spurious peak values ​​caused by data noise. The combined use of these two conditions significantly improves the accuracy and recall of peak detection. In the flexible window optimization, a length adjustment coefficient is introduced... For window width Penalize the candidate window in the neighborhood. Local competitive selection is performed within the window, allowing the window boundary to be adaptively determined by multi-scale competition rather than by hard segmentation with a fixed threshold. When the increase in information density of a candidate window is insufficient to offset the data cost caused by its increased length, the window will be automatically eliminated, thus achieving an optimal balance between window information density, compact length, and tolerance for reproductive period shifts.

[0014] In one possible implementation, step A6 further includes: A61. Select all candidate features for land cover categories. Maximum cumulative gain as category The strongest single-feature gain ; A62. Initialize Subsets All candidate features constitute the complete set of candidate features. ; A63. For each candidate feature Simulate its joining of a subset After, subset Category Cumulative discrimination :

[0015] in, For subset Features Category The cumulative gain; Candidate features Category The cumulative gain; A64. Calculate candidate features Pair subset Overall rating :

[0016] in, This is the continuous reward coefficient; For indicator functions; A65. Selecting a Pair of Subsets The largest candidate feature in the overall score Join a subset Then, determine whether the subset meets the termination condition. If it does, use the subset as the input feature subset; otherwise, return to step S63.

[0017] The beneficial effects of the above technical solution are as follows: This solution takes maximizing gain as its core principle, combined with temporal continuity rewards, to ensure that the selected subset of input features can fairly cover all land cover categories (especially ensuring that the discriminative power of categories with few samples is not overwhelmed by dominant categories), while also being as concentrated as possible in time series, and reducing the number of remote sensing image scenes actually needed to be acquired. Candidate features Add to current subset Later, for categories The cumulative discriminative power is determined by the cumulative gain of the selected features. With candidate feature gain The sum is the result. Overall score. Firstly The weakest class after normalization is selected from all categories to ensure that each new feature maximizes the improvement of the currently weakest class; this is then followed by a continuous reward. Under the same discriminative power, this mechanism prioritizes candidate features adjacent to the already selected features, thereby reducing the number of remote sensing image scenes corresponding to the final feature subset. This mechanism fundamentally solves the problem of traditional XGBoost feature importance ranking biasing towards dominant classes and neglecting weaker classes.

[0018] In one possible implementation, the termination condition is the satisfaction of any one of conditions one, two, and three, where condition one is that the number of selected candidate features reaches a preset upper limit. Condition two is The preset contribution threshold is reached; condition three is a newly added candidate feature pair. The increase is lower than the preset minimum gain. .

[0019] In one possible implementation, the network model is trained using a two-stream cross-modal attention network, which includes: Two Bi-LSTMs are used to extract the optical hidden state vector and SAR hidden state vector of optical and SAR features in each time phase, respectively. The cross-modal fusion layer is used to calculate the modal bias of two hidden state vectors based on the cloud probability generated by remote sensing optics; the two modal biases are subjected to a Softmax operation to obtain modal attention weights, and the two hidden state vectors are fused using the modal attention weights to obtain the cross-modal fusion features for each time phase; The temporal attention layer is used to generate temporal attention weights guided by the growth season mask based on cross-modal fusion features and the input growth season mask. The land cover classification module is used to weight the cross-modal fusion features of all time phases within the complete growth cycle of land cover according to the temporal attention weight. Then, after passing through a fully connected layer and a Softmax classifier, the land cover category probability is obtained.

[0020] In one possible implementation, the method for constructing the loss function of the two-stream cross-modal attention network includes: Construct a computational model for temporal attention weights guided by the growing season mask:

[0021] in, Let be the temporal attention weight for time phase t; and These are the scores of phases t and k after masking, respectively. It is an exponential function; The total number of phases in a complete growing season; Masking for the growing season; This is a learnable temporal attention weight vector; For transpose; The cross-modal fusion feature of phase t; Constructing physical consistency loss Computational model:

[0022] , in, and These represent the time-varying consistency weight and uncertainty estimate for phase t, respectively; For a learnable optical-SAR mapping network, a two-layer fully connected network is used, with the middle layer using the ReLU activation function; and Phases Optical and SAR hidden state vectors; Let be the posterior probability function, representing the attention weights at a given time. Under these conditions, the posterior probability that this time phase belongs to a phenological stage is predicted from intermediate features by a lightweight fully connected layer connected to a Softmax classifier. This is the label for the phenological stage at time t; The activation function ensures that the uncertainty estimate is positive. is the squared Euclidean distance, used to measure the difference between the predicted and actual values; c is the land cover category; It is a learnable linear layer; The cloud probability is generated by remote sensing optics based on time phase t; Determine the multi-class cross-entropy loss based on the class probabilities output by the Softmax classifier. Predictive tags for ground features Category probability Corresponding label, category probability The expression is: , in, and These are the weight matrix and bias vector of the Softmax classifier, respectively; For context vectors; For predicting the true labels of ground features; according to , and Construct the loss function for a two-stream cross-modal attention network. :

[0023] in, and These are the entropy regularization coefficient and the physical consistency regularization coefficient, respectively.

[0024] The beneficial effects of the above technical solution are as follows: the total loss function of this solution It consists of three components: multi-class cross-entropy loss. Entropy regularization term and physical consistency loss Among them, cross-entropy loss Measure the class probabilities output by the model With real-world feature labels The difference between them directly drives the improvement of classification accuracy. The entropy regularization term... For phase The temporal attention weights (i.e., the normalized weights of the temporal attention layer output guided by the growth season mask in the formula) ), This is the entropy regularization coefficient. This term penalizes the uniform distribution of attention weights (when...). When it tends to be uniform, the entropy (tending to the maximum), prompting The spontaneous concentration on a few time phases that contribute most to the differentiation of multiple land cover types makes the attention peaks in subsequent window scoring more prominent and the baseline lower, thus making key window extraction more stable and reliable. (Physical consistency loss) In the middle, learnable mapping networks (Using a 2-layer fully connected network, with the middle layer using the ReLU activation function) with optical hidden states Given the input, predict its corresponding SAR hidden state. And compare the predicted values ​​with the actual SAR hidden state The square of the Euclidean distance between them As a constraint, this constraint is designed with reference to the physical relationship between optical characteristic NDVI and vegetation water content (VWC-SAR) backscattering revealed by the Water Cloud Model, making the cross-modal mapping physically interpretable rather than a pure black-box fitting. Time-varying weights Calculated from the confidence level of phenological stages: when phase When the phenological stage can be determined with high confidence, To reduce consistency constraints and avoid mismatch in optical-SAR relationships across phenological stages; when the determination of the time phase is ambiguous, Strengthen consistency constraints to enhance physical guidance. Uncertainty. From learnable linear layers From optical stealth state SAR hidden state Heyun Probability The prediction was positive after activation with Softplus; however, in anomalies such as at the cloud edge where the two modalities should be physically consistent but their actual characteristics are inconsistent, Automatically increases, through the denominator Reduce the consistency constraint strength for this phase to prevent forced fitting of noise. Simultaneously, the regularization term... To prevent the model from maximizing the uncertainty of all time phases to evade consistency constraints, the model achieves a balance between classification accuracy, attention concentration, and cross-modal physical consistency through the synergistic optimization of the three loss terms mentioned above. That is, while ensuring the classification accuracy of multiple land cover types, the attention distribution is made more concentrated (which is beneficial for key window extraction), and the physical rationality and robustness of the fusion of optical and SAR observations are enhanced.

[0025] In one possible implementation, the expression for calculating the confidence level of the phase observation window is:

[0026] in, For the i-th temporal observation window Confidence level; For the i-th temporal observation window Length; Let be the temporal attention weight for time phase t; In step S5, the optical and SAR features of the substitute window are input into the network model to obtain its temporal attention weights for each phase; based on the temporal attention weights of the substitute window for each phase and the cloud probability generated based on remote sensing optics, the window score of the substitute window is calculated; the expression for the window score is:

[0027] in, Time scale The next A continuous time window Window rating; The cloud probability is generated by remote sensing optics based on time phase t, which is the percentage of cloud-covered pixels in the study area corresponding to time phase t.

[0028] Secondly, a land cover classification system based on observation window and input feature optimization is provided, which includes the following steps: The image acquisition module is used to read the stored temporal observation window and input feature subset, and acquire remote sensing optical and SAR images within the corresponding temporal observation window for the complete growing season of the study area. The temporal attention weight generation module is used to extract optical and SAR features of remote sensing optical and SAR images based on the features in the input feature subset, and input them into the trained network model to obtain the temporal attention weight for each time phase. The confidence calculation module is used to calculate the confidence of a time-phase observation window based on the attention weights of all time phases corresponding to each time-phase observation window. The judgment module is used to determine whether the confidence level of the phase observation window is greater than the preset threshold. If it is, the module will proceed to the fine category map generation module; otherwise, the module will proceed to the phase observation window update module. The phase observation window update module is used to shift the phase observation window to the left or right by one phase to obtain two alternative windows, calculate the window scores of the two alternative windows, update the phase observation window with the alternative window with the largest window score, and then enter the fine category map generation module. The fine category map generation module is used to input the optical and SAR features corresponding to all temporal observation windows into the XGBoost model to obtain fine category maps of various land features.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Based on the temporal observation window and input feature subset of the previous debugging, this scheme extracts remote sensing images and SAR images corresponding to the complete growing season of the study area. Based on the input feature subset, the feature selection is carried out. In this way, only a few remote sensing images are needed, and only a small number of representative features from a few remote sensing images are needed. This can greatly reduce the amount of data to be processed, thereby reducing the data cost. In the process of land cover classification, the low confidence window can be determined by the confidence of the temporal observation window, so as to adjust it in time to ensure that the selected features of the XGBoost model are more representative, thereby improving the accuracy of land cover classification.

[0030] (2) Physically Guided Cross-Modal Dynamic Fusion. A physical constraint fusion mechanism between optical vegetation index (NDVI) and SAR backscattering is established using a Water Cloud Model (WCM) physical prior-guided optical-to-SAR mapping network and a consistency loss function. Unlike existing simple feature stitching or hidden-layer fusion methods, this method utilizes the physical relationship chain of optical feature NDVI-vegetation water content (VWC)-SAR backscattering in WCM as a qualitative guide, through a learnable mapping network. The implicit mapping from optics to SAR is automatically learned from data without the need for precise calibration of WCM empirical parameters, combining physical interpretability with data-driven flexibility. Time-varying consistency weights are introduced. and heteroscedasticity uncertainty estimation It can adaptively adjust the consistency constraint strength at different phenological stages and automatically reduce the constraint under abnormal conditions such as cloud edges, thereby maintaining the robustness of fusion when optical data quality fluctuates.

[0031] (3) Enhanced attention and specificity index guided by growing season mask. Unlike existing purely data-driven attention mechanisms, this method embeds phenological common sense into temporal attention calculation in a lightweight manner as a "growing season mask." It only requires setting a loose growing season range as a weak prior, without needing the specific and accurate growth period of the crop, which can effectively constrain attention to search for key windows within a reasonable season and avoid attention being scattered to irrelevant phases such as the winter fallow period or after harvest. The setting of the mask does not require precise phenological nodes and has a high degree of tolerance and universality.

[0032] (4) Window-Feature Joint Optimization. Traditional methods typically execute key window extraction and feature selection as two independent stages sequentially. Once the window is determined, it remains fixed and cannot be adjusted based on subsequent feature selection results. This method forms a closed-loop optimization of window extraction-spectral clustering deduplication-feature selection-window feedback through window selection feedback and iterative refinement mechanisms. Specifically, when no features within a window are selected into the final subset, the attention weight of that window is automatically reduced and the window is re-extracted. When the iterations tend to stabilize, the window compactness, feature sparsity, and class balance can be synergistically optimized under a unified joint objective function. This mechanism ensures that the final solution achieves Pareto optimality between classification accuracy and data cost, overcoming the inherent limitations of the traditional two-stage separate architecture.

[0033] (5) Category-fair and temporally compact feature subsets. Using the maximization of the weakest class gain as the core criterion, combined with spectral clustering deduplication and temporal continuity rewards, the selected few features can fairly cover all land cover categories while being as concentrated temporally as possible, significantly reducing data acquisition, storage, and computational costs. Unlike existing XGBoost feature selection methods that directly use global gain ranking and are prone to bias towards dominant categories, this method proposes the "maximization of the weakest class gain" criterion. For each feature, its gain is decomposed and normalized by category, using the gain of the weakest category among all categories as the core criterion for evaluating the feature. This is supplemented by a temporal continuity reward (encouraging the selection of features with adjacent time phases), and a feature subset is constructed through a forward greedy search. This strategy ensures that the selected few features can fairly cover all land cover categories through complementary combinations, especially ensuring that key discriminative features of sparsely sampled categories are not submerged, while also being as concentrated temporally as possible, reducing the actual number of remote sensing images required to achieve low-cost, high-efficiency, and highly fair multivariate land cover classification. Attached Figure Description

[0034] Figure 1 This is a flowchart of a land cover classification method optimized based on observation window and input features.

[0035] Figure 2 This is a schematic diagram of the principle of a land cover classification system optimized based on observation window and input features. Detailed Implementation

[0036] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0037] refer to Figure 1 , Figure 1 A flowchart of a land cover classification method based on observation window and input feature optimization is shown; for example... Figure 1 As shown, the method S includes steps S1 to S6.

[0038] In step S1, the stored temporal observation window and input feature subset are read, and remote sensing optical and SAR images within the temporal observation window corresponding to the complete growing season of the study area are obtained; in this scheme, the optical image can be Sentinel-2, and the SAR image can be Sentinel-1.

[0039] In step S2, based on the features in the input feature subset, optical and SAR features of remote sensing optical and SAR images are extracted and input into the trained network model to obtain the temporal attention weights for each time phase; the network model is trained using a two-stream cross-modal attention network.

[0040] In step S3, the confidence level of each temporal observation window is calculated based on all temporal attention weights corresponding to each temporal observation window. The expression for this confidence level is as follows:

[0041] in, For the i-th temporal observation window Confidence level; For the i-th temporal observation window Length; Let t be the temporal attention weight for time phase t.

[0042] In step S4, it is determined whether the confidence level of the phase observation window is greater than a preset threshold. If so, proceed to step S6; otherwise, proceed to step S5. The preset threshold is 0.5.

[0043] In step S5, the phase observation window is shifted left and right by at least one phase to obtain two alternative windows. The window scores of the two alternative windows are calculated. The phase observation window is updated using the alternative window with the largest window score, and then the process proceeds to step S6. In step S5, preferably shifting one time phase to the left and right, the optical and SAR features of the substitute window are input into the network model to obtain its temporal attention weight for each time phase; based on the temporal attention weight of the substitute window for each time phase and the cloud probability generated based on remote sensing optics, the window score of the substitute window is calculated; the expression for the window score is:

[0044] in, Time scale The next A continuous time window Window rating; The cloud probability is generated by remote sensing optics based on time phase t, which is the percentage of cloud-covered pixels in the study area corresponding to time phase t.

[0045] In step S6, the optical and SAR features corresponding to all temporal observation windows are input into the XGBoost model to obtain a detailed category map of various land features. After the XGBoost model is trained, it has a dual function: first, it outputs the importance score of each feature; second, it serves as the final land feature classifier, directly performing fast land feature classification on optical and SAR features.

[0046] In one embodiment of the present invention, the method for constructing the stored temporal observation window and the input feature subset includes: A1. Obtain multi-temporal remote sensing optical and SAR images of ground features in the target area over several years, covering the entire growing season. Extract the optical and SAR features of each ground feature sample and combine them with the ground feature labels of the ground features to form a dual-modal time series sample set.

[0047] Taking into account the availability of imagery in cloudy and rainy areas, crop growth rhythms, and effective coverage after optical and SAR synergy, this scheme synthesizes images at temporal resolutions of five days, one week, or one ten-day period. For multiple images within the same temporal window, the median is used for synthesis; for ten-day periods without images, linear interpolation between adjacent ten-day periods is used to fill the gap.

[0048] In step A1, raw bands (blue, green, red, red-edge, near-infrared, shortwave infrared, etc.) and general remote sensing indices (Enhanced Vegetation Index EVI, Normalized Difference Vegetation Index NDVI, Surface Water Index LSWI, etc.) are extracted from the optical images on a time-by-time basis. For specific crop types to be identified, additional phenologically specific indices with discriminative power are calculated. For example, the Yellow Flower Index (YFI) is calculated for rapeseed during its flowering period; and the Yellowness Index (YI) is calculated for crops such as corn and soybeans during their maturity and harvest period. The addition of these specific indices provides strong discriminative signals for specific crops at key time phases.

[0049] For SAR imagery, the VV and VH polarization backscattering coefficients are obtained, radiometrically calibrated and topographically corrected, and then incorporated into the feature set. Based on this, the ratio of VV to VH is calculated. Radar vegetation index (RVI), polarization ratio (PR), normalized differential polarization index (NDPI), and radar vegetation index (RVI) are included as additional SAR features.

[0050] Simultaneously, using the cloud mask information inherent in optical images (such as the SCL layer of Sentinel-2), the percentage of cloud-covered pixels within each temporal study area is calculated as the cloud probability. This is used to indicate the reliability of optical data in subsequent cross-modal fusion. For large areas spanning multiple agricultural resource and environmental zones, each zone is one-hot encoded and concatenated as a spatial feature into the feature vector of each temporal phase.

[0051] Ultimately, this forms the total three-dimensional tensor. (dimension) ),in To train the number of classification samples for ground features, For the time phase number, For the total number of features, It consists of optical characteristic numbers (including original bands, general indices, and phenological specific indices). These are SAR characteristic numbers (including backscattering coefficients and derived indices). This is for auxiliary feature numbers (i.e., partition coding). It corresponds to multiple types of land cover labels. The number of categories is denoted as ( ).

[0052] A2. Divide the dual-modal temporal sample set into a training set and a test set, and use the training set to train the dual-stream cross-modal attention network to obtain a network model that extracts the temporal attention weights for each phase.

[0053] In implementation, this scheme preferably includes two Bi-LSTMs, a cross-modal fusion layer, a temporal attention layer, and a ground feature classification module.

[0054] Two Bi-LSTMs are used to extract the optical hidden state vector and SAR hidden state vector of optical and SAR features at each time phase, respectively; the input data of one of the Bi-LSTMs is the optical flow consisting of the number of optical features plus auxiliary features. The other input data for Bi-LSTM is a SAR stream consisting of SAR feature numbers and auxiliary feature numbers. .

[0055] The two expressions for extracting the temporal hidden state using Bi-LSTM are as follows:

[0056] in, and Phases The optical and SAR hidden state vectors, whose dimensions are determined by the number of Bi-LSTM hidden units. Decision (suggested value: 64-256). , Two Bi-LSTM networks with independent parameters are used to learn the dependencies within the optical and SAR time series, respectively, providing a mode-specific time series representation for cross-modal fusion.

[0057] The cross-modal fusion layer is used to calculate the modal bias of two hidden state vectors based on the cloud probability generated by remote sensing optics. A softmax operation is performed on the two modal biases to obtain modal attention weights, which are then used to fuse the two hidden state vectors to obtain the cross-modal fusion features for each time phase. The detailed implementation process of the cross-modal fusion layer is as follows: cloud probability As a priori signal, through a learnable shared transformation matrix (dimension) , (For the attention hidden layer dimension, a value of 32-128 is recommended) and it is mapped to modal bias:

[0058]

[0059] In the formula, and For mode-specific learnable transformation matrices (dimension) This maps the hidden states to the attention scoring space. Scoring vectors for learnable attention (dimensions) The transformed features are mapped to scalar scores. and (Unnormalized modality importance score).

[0060] Modal attention weights are obtained using Softmax:

[0061] In the formula, and The optical and SAR modal attention weights for time phase t, i.e., the contribution ratios, satisfy the following conditions: .when hour, The model depends on optics; when hour, The model automatically switches to SAR.

[0062] The fusion feature is a weighted sum of the hidden states of the two modalities:

[0063] In the formula, For the first The cross-modal fusion characteristics of time phase.

[0064] The temporal attention layer is used to generate temporal attention weights guided by the growth season mask based on cross-modal fusion features and the input growth season mask. Its detailed implementation process is as follows: The growing season mask is used to constrain the temporal attention. Based on the common sowing and harvesting periods of the main crops in the study area, a loose growing season range is set (for example, October to June of the following year for overwintering crops and May to October for autumn-harvested crops). The temporal weights outside this range are forcibly suppressed, so that the model only looks for key windows within the growing season and avoids the attention being diverted to invalid temporal phases outside the growing season.

[0065] Define the growing season mask If the first If the time phase falls within the preset growing season range, then ;otherwise The range of the growing season can be determined by consulting local agricultural zoning data or analyzing the multi-year average NDVI time series curve (the period when NDVI is consistently above 0.2 is taken as the growing season). This range has considerable tolerance and does not require precise phenological nodes.

[0066] In the calculation of temporal attention, a mask is applied to the original score to zero out the weights of temporal phases outside the growing season:

[0067] In the formula, For learnable temporal attention weight vectors (dimensions) ), which maps the fused features to scalar scores; For the first The fusion characteristics of time phases; The score after masking; The normalized attention weight for time phase t is defined as follows: a larger value indicates that the attention weight is more critical for multi-class differentiation at that time phase, and satisfies the following conditions: .for phase of time, Its normalized weight Although not absolute zero, it is negligible compared to other high-scoring phases, effectively excluding them from the critical window search. The growing season mask only filters out obviously irrelevant phases; subsequent flexible window extraction steps will further narrow down to the truly critical sub-intervals.

[0068] The land cover classification module is used to weight the cross-modal fusion features of all time phases within the complete growth cycle of land cover according to the temporal attention weight. Then, after passing through a fully connected layer and a Softmax classifier, the land cover category probability is obtained.

[0069] In this scheme, the dual-stream Bi-LSTM uses optical feature tensors respectively. and SAR feature tensor As input, the temporal dependencies of each mode are captured through independent gating mechanisms, and the optical hidden states are output. and SAR hidden state This provides modality-specific temporal representations for subsequent fusion. The cross-modal fusion layer uses cloud probability... Given a priori signals, transformed by a learnable transformation matrix After mapping to modal biases, the results are added to the transformations of the two hidden states, and then processed by the hyperbolic tangent and the attention scoring vector. Obtain the unnormalized score and Then, the values ​​are normalized using Softmax to become modal attention weights. and ,satisfy When cloud probability hour The model tends to be optical; when hour The model automatically switches to SAR, thereby fusing features. This achieves dynamic switching between "optical analysis for clear skies and SAR analysis for cloudy / rainy skies." The temporal attention layer utilizes a growing season mask. ( This indicates the time phase within the growing season. (Indicates the original score outside the growing season) Masking is performed to effectively suppress temporal attention weights outside the growing season, confining the attention search space to a reasonable seasonal range and avoiding interference from irrelevant temporal phases. The land cover classification module uses the temporally attention-weighted context vector. Input, processed by a fully connected layer and a Softmax classifier, outputting the probability of land cover categories. This forms an end-to-end forward propagation path. The modules in the entire network structure have clearly defined responsibilities and work collaboratively, ensuring classification accuracy while possessing strong physical interpretability.

[0070] In one embodiment of the present invention, the method for constructing the loss function of the two-stream cross-modal attention network includes: Construct a computational model for temporal attention weights guided by the growing season mask:

[0071] in, Let be the temporal attention weight for time phase t; and These are the scores of phases t and k after masking, respectively. It is an exponential function; The total number of phases in a complete growing season; Masking for the growing season; This is a learnable temporal attention weight vector; For transpose; The cross-modal fusion feature of phase t; Constructing physical consistency loss Computational model:

[0072] , in, and These represent the time-varying consistency weight and uncertainty estimate for phase t, respectively; For a learnable optical-SAR mapping network, a two-layer fully connected network is used, with the middle layer using the ReLU activation function; and Phases Optical and SAR hidden state vectors; Let be the posterior probability function, representing the attention weights at a given time. Under these conditions, the posterior probability that this time phase belongs to a phenological stage is predicted from intermediate features by a lightweight fully connected layer connected to a Softmax classifier. This is the label for the phenological stage at time t; The activation function ensures that the uncertainty estimate is positive. is the squared Euclidean distance, used to measure the difference between the predicted and actual values; c is the land cover category; It is a learnable linear layer; The cloud probability is generated by remote sensing optics based on time phase t; Determine the multi-class cross-entropy loss based on the class probabilities output by the Softmax classifier. Predictive tags for ground features Category probability Corresponding label, category probability The expression is: , in, and These are the weight matrix and bias vector of the Softmax classifier, respectively; For context vectors; For predicting the true labels of ground features; according to , and Construct the loss function for a two-stream cross-modal attention network. :

[0073] in, and These are the entropy regularization coefficient and the physical consistency regularization coefficient, respectively.

[0074] A3. Input the samples from the test set into the network model to obtain the temporal attention weights for each sample at each time phase. Based on the temporal attention weights and cloud probabilities, determine several temporal observation windows. The methods for determining the temporal observation windows include: A31. Averaging the phase attention weights for the same time phase yields the average phase attention vector for all time phases. ( The average cloud probability for all samples in the test set at the same time phase is calculated. .

[0075] A32. Obtaining a Timescale Set ,like For each time scale, the system slides across the full-temporal sequence of the complete growing season with a preset step size. This scheme preferably slides with a step size of 1 to generate all possible continuous temporal windows. ,in For window number, This is the initial phase.

[0076] A33. Calculate the window score for each consecutive temporal phase window based on all average temporal attention vectors and average cloud probabilities within the consecutive temporal phase windows:

[0077] In the formula, As a quality weight, even if the attention is high, the effective score of the phases polluted by clouds is weakened, and the window is automatically biased towards the clear phases; The higher the value, the greater the information density within the window and the better the data availability.

[0078] A34. Select the average temporal attention weights from the average temporal attention vector that meet the preset conditions, and mark their corresponding temporal phases as core temporal phases. And generate the neighborhood of each core temporal phase; wherein, the preset condition is that the average temporal attention simultaneously satisfies condition A and condition B: condition A is and , , and Phases , and Average temporal attention; condition B is Higher than the neighborhood The mean of the average temporal attention plus 1.5 times the standard deviation of the neighborhood. Set the neighborhood expansion radius. (Extend by one ten-day period to the left and right of the corresponding peak, i.e., ±10 days), define the neighborhood range. .

[0079] A35. Filter out the phase windows located in the neighborhood among all continuous phase windows, and select the phase observation window with the highest window score in each neighborhood as the phase observation window.

[0080] The expression for selecting the time series window with the highest window score as the time phase observation window within the same neighborhood is:

[0081] in, For the i-th neighborhood The selected timing window; For the i-th core phase The neighborhood; Time scale The next A continuous time window Window rating; This is a length adjustment factor (recommended value 0.01-0.05, typical value 0.02). This is a penalty item based on the scale.

[0082] A4. For samples in a dual-modal temporal sample set (i.e., the original three-dimensional tensor of the samples) The time phases and all optical and SAR features of the data within all time phase observation windows are preserved and flattened into a two-dimensional table. , dimension ,in , The number of samples; The total number of features for each time phase, including It consists of optical characteristic numbers (including original bands, general indices, and phenological specific indices). These are SAR characteristic numbers (including backscattering coefficients and derived indices). This is for auxiliary feature numbers (i.e., partition coding).

[0083] The retained features are deduplicated using spectral clustering to obtain multiple candidate features; the detailed implementation process of this step is as follows: All within the window Perform spectral clustering on the features. Calculate the Spearman correlation coefficient matrix between the features. Construct a similarity map and use spectral clustering to classify features. Clusters ( (Automatically determined by spectral gaps). Within each cluster, only the representative feature with the highest average correlation to other features in that cluster is retained, while the rest are temporarily masked. After this deduplication, the number of candidate features is reduced from... Down to Furthermore, it retains various representative feature types, thus obtaining multiple candidate features.

[0084] A5. Input all candidate features of all samples in the bimodal time-series sample set into the XGBoost model, and output the cumulative gain of each candidate feature for the land cover category; the XGBoost model here is the same as the XGBoost model in step S5, both based on... (Only includes data retained after deduplication by S4.1 spectral clustering) (representative features) and multiple labels The trained XGBoost ensemble learning model.

[0085] To train an initial XGBoost model, you can configure the following: the number of trees. Maximum depth max_depth = 6, learning rate Regularization parameters , The subsampling ratio is 0.8, the feature sampling ratio is 0.8, and the objective function is multi-class cross-entropy. Single-tree regularization is as follows: .

[0086] After training, for each feature (Corresponding to a specific "phase-band / exponential / scattering" combination, such as "NDVI of the 12th phase"), its statistics are calculated across all Gain when selected as a split node in a tree (i.e., the reduction in loss function before and after splitting), and sorted by the proportion of each category of samples within the node. The gain is assigned to the corresponding category to obtain the feature. Category Cumulative gain:

[0087] In the formula, Category in the node The proportion of the number of samples to the total number of samples in the nodes; "split on inclass "Indicates features" For a given split, which is a splitting node, its gain is calculated as follows: The proportion is included in the category The account. Same characteristics It may be repeatedly selected as the basis for splitting at multiple different nodes in multiple trees, and the gains of these splitting events are all accumulated.

[0088] The XGBoost model is saved after training and used in the following two stages: (1) based on its output. Perform class-balanced forward feature selection to obtain the minimum optimal subset of input features. (2) Directly used as a classifier, input according to The filtered features are used to output classification results for multiple types of land features.

[0089] A6. Based on the cumulative gain, each time the candidate feature with the largest comprehensive score for the subset is selected and added to the subset, and when the subset meets the termination condition, the subset is used as the input feature subset. A7. Determine whether at least one feature is selected into the feature subset within the time phase observation window or whether the time phase observation window has not changed in the last two times. If so, retain the most recent time phase observation window and the feature subset; otherwise, proceed to step A8. A8. Reduce the temporal attention weight of the window without features by a preset ratio (i.e., the weight corresponding to that window). Multiply by the attenuation factor After resetting the phase observation window, return to step A4.

[0090] In one embodiment of the present invention, step A6 further includes: A61. Select all candidate features for land cover categories. Maximum cumulative gain as category The strongest single-feature gain .

[0091] A62. Initialize Subsets All candidate features constitute the complete set of candidate features. ; A63. For each candidate feature Simulate its joining of a subset After, subset Category Cumulative discrimination :

[0092] in, For subset Features Category The cumulative gain; Candidate features Category The cumulative gain; A64. Calculate candidate features Pair subset Overall rating :

[0093] in, This is the continuous reward coefficient (range 0.02-0.10, typical value 0.05). As an indicator function, when candidate features Corresponding time phase and selected subset The value is 1 if at least one feature is in a temporally adjacent phase (i.e., the phase index difference is 1), otherwise it is 0. This mechanism encourages the selection of features that are temporally continuous, reducing the number of image scenes required.

[0094] A65. Selecting a Pair of Subsets The largest candidate feature in the overall score Join a subset Select the candidate feature with the highest overall score. The expression is: .

[0095] Next, determine whether the subset meets the termination condition. If it does, use the subset as the input feature subset; otherwise, return to step S63.

[0096] In the process of constructing the temporal observation window and the input feature subset, this scheme uses temporal attention weights. With cloud probability Combined with window scoring, this achieves a balance between window information density and data availability. Specifically, window scoring... middle, For all test samples in the first The average temporal attention weight of a time phase reflects the importance of that time phase in distinguishing multiple land cover types; For all test samples in the first Average cloud probability of time phase The effective contribution of cloud-contaminated phases is suppressed as a quality weight. The statistical caliber is consistent after averaging the two phases, ensuring matching of the two factor dimensions in the scoring formula. Flexible window optimization is employed. middle, This is the length adjustment factor. This is a scaling penalty term. This formula ensures that a longer window is selected only when the information gain from window lengthening exceeds the penalty increment, thus avoiding data redundancy caused by excessively wide windows. The overall score during the feature selection stage... middle, Candidate features Add to current subset Later, for categories Cumulative discrimination power For category The strongest single-feature gain benchmark, The weakest discriminative power among all categories after normalization is selected as the core evaluation. For continuous reward coefficient, As an indicator function, extra points are awarded when the time phase of a candidate feature is adjacent to the time phase of a selected feature. This design enables the final constructed time phase observation window and the subset of input features to achieve Pareto optimality between classification accuracy and data cost.

[0097] In implementation, this scheme preferably terminates the program under any one of conditions one, two, or three, where condition one is that the number of selected candidate features reaches a preset upper limit. Condition two is The preset contribution threshold is reached; condition three is a newly added candidate feature pair. The increase is lower than the preset minimum gain. .

[0098] When optimizing the temporal observation window, this scheme unifies window compactness, feature sparsity, and class balance into a single joint optimization objective:

[0099] In the formula, the first term represents the weakest class discrimination power. This is the characteristic sparsity penalty coefficient (with a value of 0.1). This is the window compactness penalty coefficient (value 0.05). The iterative refinement process essentially optimizes this joint objective, aiming to achieve Pareto optimality between classification accuracy and data cost in the final solution.

[0100] like Figure 2 As shown, this scheme also provides a land cover classification system based on observation window and input feature optimization, which includes the following steps: The image acquisition module is used to read the stored temporal observation window and input feature subset, and acquire remote sensing optical and SAR images within the corresponding temporal observation window for the complete growing season of the study area. The temporal attention weight generation module is used to extract optical and SAR features of remote sensing optical and SAR images based on the features in the input feature subset, and input them into the trained network model to obtain the temporal attention weight for each time phase. The confidence calculation module is used to calculate the confidence of a time-phase observation window based on the attention weights of all time phases corresponding to each time-phase observation window. The judgment module is used to determine whether the confidence level of the phase observation window is greater than the preset threshold. If it is, the module will proceed to the fine category map generation module; otherwise, the module will proceed to the phase observation window update module. The phase observation window update module is used to shift the phase observation window to the left or right by one phase to obtain two alternative windows, calculate the window scores of the two alternative windows, update the phase observation window with the alternative window with the largest window score, and then enter the fine category map generation module. The fine category map generation module is used to input the optical and SAR features corresponding to all temporal observation windows into the XGBoost model to obtain fine category maps of various land features.

[0101] In summary, the innovation of this solution lies in the first-time organic integration of the window discovery capability of the temporal attention mechanism with the feature precision selection capability of ensemble learning (XGBoost model), and it is specifically optimized for complex multivariate classification scenarios, effectively serving the goal of lightweight investment in large-scale land cover classification, thereby reducing data costs and improving classification accuracy.

Claims

1. A land cover classification method based on observation window and input feature optimization, characterized in that, Including the following steps: S1. Read the stored temporal observation window and input feature subset, and obtain remote sensing optical and SAR images within the temporal observation window corresponding to the complete growing season of the study area; S2. Based on the features in the input feature subset, extract the optical and SAR features of remote sensing optical and SAR images, and input them into the trained network model to obtain the temporal attention weights for each time phase. S3. Calculate the confidence level of each phase observation window based on all phase attention weights corresponding to each phase observation window. S4. Determine whether the confidence level of the phase observation window is greater than the preset threshold. If yes, proceed to step S6; otherwise, proceed to step S5. S5. Shift the phase observation window to the left or right by at least one phase to obtain two alternative windows. Calculate the window scores of the two alternative windows. Update the phase observation window with the alternative window with the largest window score and proceed to step S6. S6. Input the optical and SAR features corresponding to all temporal observation windows into the XGBoost model to obtain detailed category maps of various land features.

2. The land cover classification method according to claim 1, characterized in that, Methods for constructing stored temporal observation windows and input feature subsets include: A1. Acquire multi-temporal remote sensing optical and SAR images of ground features in the target area over several years, covering the complete growing season. Extract the optical and SAR features of each ground feature sample and combine them with the ground feature labels of the ground features to form a dual-modal time series sample set. A2. Divide the dual-modal temporal sample set into a training set and a test set, and use the training set to train the dual-stream cross-modal attention network to obtain a network model that extracts the temporal attention weights for each phase. A3. Input the samples in the test set into the network model to obtain the temporal attention weights for each sample at each time phase, and determine several temporal observation windows based on the temporal attention weights and cloud probabilities; A4. For samples in the dual-modal time series sample set, retain the temporal phase and all optical and SAR features that are within a defined observation window, and perform spectral clustering to remove duplicates from the retained features to obtain multiple candidate features. A5. Input all candidate features from the dual-modal time series sample set into the XGBoost model and output the cumulative gain of each candidate feature for the land cover category; A6. Based on the cumulative gain, each time the candidate feature with the largest comprehensive score for the subset is selected and added to the subset, and when the subset meets the termination condition, the subset is used as the input feature subset. A7. Determine whether at least one feature is selected into the feature subset within the time phase observation window or whether the time phase observation window has not changed in the last two times. If so, retain the most recent time phase observation window and the feature subset; otherwise, proceed to step A8. A8. Reduce the temporal attention weight of the window without features by a preset ratio, and return to step A4 after redetermining the temporal observation window.

3. The land cover classification method according to claim 2, characterized in that, Methods for determining the temporal observation window based on temporal attention weights and cloud probabilities include: A31. Average the phase attention weights for the same phase to obtain the average phase attention vector for all phases; average the cloud probabilities for the same phase for all samples in the test set to obtain the average cloud probabilities for all phases. A32. Obtain the set of time scales. For each time scale, slide across the full-phase sequence of the complete growing season with a preset step size to generate all possible continuous phase windows. A33. Calculate the window score for each consecutive time phase window based on all average time attention vectors and average cloud probabilities within the consecutive time phase window. A34. Select the average temporal attention weights that meet the preset conditions from the average temporal attention vectors, mark their corresponding temporal phases as core temporal phases, and generate the neighborhood of each core temporal phase; A35. Filter out the phase windows located in the neighborhood among all continuous phase windows, and select the phase observation window with the highest window score in each neighborhood as the phase observation window.

4. The land cover classification method according to claim 3, characterized in that, The preset condition is that the average temporal attention simultaneously satisfies both condition A and condition B: Condition A is and , , and Phases , and Average temporal attention; condition B is Higher than the neighborhood The mean of the average temporal attention plus 1.5 times the standard deviation of the neighborhood; The expression for selecting the time series window with the highest window score as the time phase observation window within the same neighborhood is: in, For the i-th neighborhood The selected timing window; For the i-th core phase The neighborhood; Time scale The next A continuous time window Window rating; This is the length adjustment factor; This is a penalty item based on the scale.

5. The land cover classification method according to claim 2, characterized in that, Step A6 further includes: A61. Select all candidate features for land cover categories. Maximum cumulative gain as category The strongest single-feature gain ; A62. Initialize Subsets All candidate features constitute the complete set of candidate features. ; A63. For each candidate feature Simulate its joining of a subset After, subset Category Cumulative discrimination : in, For subset Features Category The cumulative gain; Candidate features Category The cumulative gain; A64. Calculate candidate features Pair subset Overall rating : in, This is the continuous reward coefficient; For indicator functions; A65. Selecting a Pair of Subsets The largest candidate feature in the overall score Join a subset Then, determine whether the subset meets the termination condition. If it does, use the subset as the input feature subset; otherwise, return to step S63.

6. The land cover classification method according to claim 5, characterized in that, The termination condition is that any one of conditions one, two, or three is met, where condition one is that the number of selected candidate features reaches a preset upper limit. Condition two is The preset contribution threshold is reached; condition three is a newly added candidate feature pair. The increase is lower than the preset minimum gain. .

7. The method for classifying land features according to claim 1, characterized in that, The network model is trained using a two-stream cross-modal attention network, which includes: Two Bi-LSTMs are used to extract the optical hidden state vector and SAR hidden state vector of optical and SAR features in each time phase, respectively. The cross-modal fusion layer is used to calculate the modal bias of two hidden state vectors based on the cloud probability generated by remote sensing optics; the two modal biases are subjected to a Softmax operation to obtain modal attention weights, and the two hidden state vectors are fused using the modal attention weights to obtain the cross-modal fusion features for each time phase; The temporal attention layer is used to generate temporal attention weights guided by the growth season mask based on cross-modal fusion features and the input growth season mask. The land cover classification module is used to weight the cross-modal fusion features of all time phases within the complete growth cycle of land cover according to the temporal attention weight. Then, after passing through a fully connected layer and a Softmax classifier, the land cover category probability is obtained.

8. The method for classifying land features according to claim 7, characterized in that, The method for constructing the loss function of the dual-stream cross-modal attention network includes: Construct a computational model for temporal attention weights guided by the growing season mask: in, Let be the temporal attention weight for time phase t; and These are the scores of phases t and k after masking, respectively. It is an exponential function; The total number of phases in a complete growing season; Masking for the growing season; This is a learnable temporal attention weight vector; For transpose; The cross-modal fusion feature of phase t; Constructing physical consistency loss Computational model: ， in, and These represent the time-varying consistency weight and uncertainty estimate for phase t, respectively; For a learnable optical-SAR mapping network, a two-layer fully connected network is used, with the middle layer using the ReLU activation function; and Phases Optical and SAR hidden state vectors; Let be the posterior probability function, representing the attention weights at a given time. Under these conditions, the posterior probability that this time phase belongs to a phenological stage is predicted from intermediate features by a lightweight fully connected layer connected to a Softmax classifier. This is the label for the phenological stage at time t; For activation functions; c is the squared Euclidean distance; c is the land feature category; It is a learnable linear layer; The cloud probability is generated by remote sensing optics based on time phase t; Determine the multi-class cross-entropy loss based on the class probabilities output by the Softmax classifier. Predictive tags for ground features Category probability Corresponding label, category probability The expression is: ， in, and These are the weight matrix and bias vector of the Softmax classifier, respectively; For context vectors; For predicting the true labels of ground features; according to , and Construct the loss function for a two-stream cross-modal attention network. : in, and These are the entropy regularization coefficient and the physical consistency regularization coefficient, respectively.

9. The method for classifying land features according to claim 1, characterized in that, The expression for calculating the confidence level of the phase observation window is: in, For the i-th temporal observation window Confidence level; For the i-th temporal observation window Length; Let be the temporal attention weight for time phase t; In step S5, the optical and SAR features of the substitute window are input into the network model to obtain its temporal attention weights for each phase; based on the temporal attention weights of the substitute window for each phase and the cloud probability generated based on remote sensing optics, the window score of the substitute window is calculated; the expression for the window score is: in, Time scale The next A continuous time window Window rating; The cloud probability is generated by remote sensing optics based on time phase t, which is the percentage of cloud-covered pixels in the study area corresponding to time phase t.

10. A land cover classification system based on observation window and input feature optimization, characterized in that, Including the following steps: The image acquisition module is used to read the stored temporal observation window and input feature subset, and acquire remote sensing optical and SAR images within the corresponding temporal observation window for the complete growing season of the study area. The temporal attention weight generation module is used to extract optical and SAR features of remote sensing optical and SAR images based on the features in the input feature subset, and input them into the trained network model to obtain the temporal attention weight for each time phase. The confidence calculation module is used to calculate the confidence of a time-phase observation window based on the attention weights of all time phases corresponding to each time-phase observation window. The judgment module is used to determine whether the confidence level of the phase observation window is greater than the preset threshold. If it is, the module will proceed to the fine category map generation module; otherwise, the module will proceed to the phase observation window update module. The phase observation window update module is used to shift the phase observation window to the left or right by one phase to obtain two alternative windows, calculate the window scores of the two alternative windows, update the phase observation window with the alternative window with the largest window score, and then enter the fine category map generation module. The fine category map generation module is used to input the optical and SAR features corresponding to all temporal observation windows into the XGBoost model to obtain fine category maps of various land features.

Images