Crop classification method based on multi-source satellite images

Abstract

The present invention relates to the field of satellite remote sensing application technology and discloses a crop classification method based on multi-source satellite imagery. The method first acquires multi-temporal satellite imagery data, extracts spectral reflectance feature generation matrices, and constructs a classification model set. A feature parameter category set is set, based on which texture, vegetation index parameters, and spatial resolution data of historical images are acquired to establish a feature fusion association model. Spectral and texture feature parameters of the crop area in the current image are extracted and combined with model optimization to generate an optimized feature set. Classification logic is updated based on the optimized feature set and classification model, a target classification scheme is generated, and spatial relationships are calibrated. Finally, crop growth cycle data is acquired and temporal feature association rules are established to optimize the spatial distribution map of crop types. Through multi-source feature fusion and dynamic optimization, this method improves the accuracy and reliability of crop classification and is suitable for precision agriculture management scenarios.

Description

Technical Field

[0001] The present invention relates to the field of satellite remote sensing application technology, and in particular to a crop classification method based on multi-source satellite images. Background Art

[0002] With the acceleration of agricultural modernization, precision agriculture management is increasingly demanding the rapid and accurate identification of crop types. Traditional crop classification methods, which rely primarily on ground surveys or single-sensor data, suffer from limited coverage, poor timeliness, high labor costs, and significant weather impacts. These methods struggle to meet the needs of large-scale, dynamic monitoring.

[0003] Satellite remote sensing technology, due to its advantages in macroscopic, dynamic, rapid, and multi-temporal observation, has become an important means of crop classification. However, existing crop classification technologies based on satellite imagery still face numerous challenges. For one thing, the information acquired by a single satellite sensor is limited. When classification is performed solely based on spectral features, it is susceptible to interference from the phenomenon of identical objects with different spectra, and different objects with the same spectra, resulting in insufficient classification accuracy. For example, the same crop at different growth stages may have different spectral reflectances, while certain different types of crops may exhibit similar spectral characteristics in specific bands, causing classification confusion.

[0004] On the other hand, existing methods do not fully utilize the fusion of multi-source satellite imagery data. Multi-source satellite imagery contains information from different bands (such as near-infrared, red edge, and shortwave infrared), texture features, vegetation indices, and spatial resolution data. Effectively integrating these multi-dimensional features to build an accurate classification model is a key issue in improving crop classification accuracy. Furthermore, when processing multi-temporal data, traditional classification models lack effective utilization of the temporal characteristics of crop growth cycles, making it difficult to dynamically reflect crop growth changes. This results in insufficient temporal continuity and reliability in classification results.

[0005] Furthermore, existing technologies often use fixed feature selection algorithms and static model parameter settings during feature extraction and model training. These algorithms are unable to adapt to the characteristics of image data from different regions and time periods, resulting in poor model generalization. For example, land cover types vary greatly across regions, and factors such as soil background and climatic conditions can affect the spectral and textural characteristics of crops. Fixed model parameters make it difficult to achieve ideal classification results across a wide range of complex environments.

[0006] Furthermore, satellite imagery can be affected by atmospheric disturbances, sensor noise, and other factors during acquisition and transmission, leading to geometric distortion and radiometric errors, which in turn affect the accuracy of feature extraction and the reliability of classification results. Existing geometric calibration methods are mostly based on fixed transformation models, which fail to fully account for the dynamic changes in satellite positioning accuracy and image distortion coefficients, and calibration accuracy needs to be improved. Summary of the Invention

[0007] The purpose of the present invention is to provide a crop classification method based on multi-source satellite images to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a crop classification method based on multi-source satellite images, the method comprising:

[0009] S1. Acquire multi-temporal satellite image data, extract spectral reflectance characteristics of different bands, and generate a spectral feature matrix;

[0010] Construct a set of crop classification models based on the spectral feature matrix;

[0011] S2. setting characteristic parameter categories corresponding to a plurality of land cover types to form a characteristic parameter category set;

[0012] Based on the feature parameter category set, texture feature parameters, vegetation index parameters and spatial resolution data from multiple sets of historical image data are obtained, and a feature fusion association model is established;

[0013] S3, extracting spectral characteristic parameters and texture characteristic parameters of each crop area in the current satellite image according to the feature parameter category set, performing feature optimization in combination with the feature fusion association model, and generating an optimized spectral feature set and an optimized texture feature set;

[0014] S4. Update the current image classification logic based on the optimized spectral feature set, optimized texture feature set and crop classification model set to generate a target classification scheme; calibrate the spatial relationship of the feature extraction link in the target classification scheme.

[0015] Preferably, said S1 comprises the following steps:

[0016] S11. Define multiple crop types to form a crop type set; collect corresponding near-infrared band, red-edge band, and short-wave infrared band data according to the crop type set to generate a spectral feature matrix;

[0017] S12. Input the spectral feature matrix into the classification model generation module, and construct a crop classification model set through a preset feature selection algorithm.

[0018] Preferably, said S2 comprises the following steps:

[0019] S21. Setting multiple feature parameter categories, including a normalized vegetation index threshold, a spatial texture complexity coefficient, and a surface temperature range, to form a feature parameter category set;

[0020] Based on the crop type set and characteristic parameter category set, historical texture feature parameters, vegetation index combinations and corresponding crop recognition accuracy data are obtained to construct historical feature matrix, historical vegetation matrix and historical accuracy dataset;

[0021] S22. Use the historical feature matrix, the historical vegetation matrix, and the historical precision data set to train an initial feature fusion association model to obtain a feature fusion association model.

[0022] Preferably, the S22 includes the following steps:

[0023] S221, establishing an initial feature fusion association model based on the crop type set and the feature parameter category set;

[0024] S222, inputting the historical feature matrix and the historical vegetation matrix into the initial feature fusion association model row by row to generate a historical accuracy prediction data set;

[0025] S223. Setting an accuracy error threshold; calculating a cross entropy error between the historical accuracy prediction dataset and the historical accuracy dataset;

[0026] When the cross entropy error exceeds the precision error threshold, the gradient descent algorithm is used to iteratively correct the hyperparameters of the initial feature fusion association model until the error is lower than the error threshold, thus forming a feature fusion association model.

[0027] Preferably, the hyperparameter correction process in S223 adopts a dynamic learning rate adjustment mechanism.

[0028] Preferably, said S3 comprises the following steps:

[0029] S31, obtaining actual spectral characteristic parameters and texture characteristic parameters of each block area in the current satellite image, and generating a real-time spectrum set and a real-time texture set;

[0030] S32, setting a classification accuracy target value; inputting the real-time spectrum set and the real-time texture set into the feature fusion association model to calculate the current classification accuracy prediction value;

[0031] When the current classification accuracy prediction value does not reach the target value, the real-time spectrum set and the real-time texture set are feature enhanced until the target conditions are met, and the optimized spectrum feature set and the optimized texture feature set are output.

[0032] Preferably, the feature enhancement in S32 includes the following steps:

[0033] S321. Create a feature optimization population and define the maximum evolutionary generation and the current evolutionary generation;

[0034] S322, setting the adjustment range of the spectral characteristic parameters and the texture characteristic parameters to generate an initial population position matrix;

[0035] S323, constructing a population fitness function based on the current classification accuracy prediction value;

[0036] S324, updating the population position through multiple rounds of evolution, and recalculating the fitness value and the global optimal solution after each round of evolution;

[0037] S325. When the evolutionary generation reaches the upper limit or the fitness value meets the convergence condition, the global optimal solution is extracted as the optimization result.

[0038] Preferably, said S4 comprises the following steps:

[0039] S41, loading the optimized spectral feature set and texture feature set into the image classification unit to generate a target classification scheme;

[0040] S42. Identify the spatial coordinates of the feature matching sequence in the target classification scheme, and perform geometric calibration according to a preset spatial resolution constraint.

[0041] Preferably, the geometric calibration in S42 adopts a coordinate affine transformation algorithm, and the coordinate affine transformation algorithm dynamically adjusts the feature matching parameters according to the satellite positioning accuracy and the image distortion coefficient.

[0042] Preferably, the method further comprises:

[0043] S5. Obtain crop growth cycle data of different phase images in the target classification scheme, establish temporal feature association rules; and optimize the spatial distribution map of crop types in the classification scheme based on the temporal feature association rules.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] In terms of multi-source feature fusion, this method not only extracts the spectral reflectance characteristics of different bands of satellite images, but also integrates multi-dimensional information such as texture feature parameters, vegetation index parameters and spatial resolution data. By constructing a feature fusion association model, dynamic optimization of spectral and texture features is achieved. For example, in the feature enhancement process, the use of evolutionary algorithms to iteratively optimize spectral and texture feature parameters can effectively solve the problem of different spectra for the same object and the same spectrum for different objects, significantly improving classification accuracy. Compared with traditional methods that rely solely on spectral features, the present invention can more comprehensively characterize the characteristic differences of crops through the synergistic effect of multi-source features, thereby improving classification accuracy.

[0046] In terms of model construction and training, by defining multiple crop types and collecting multi-band data such as near-infrared, red-edge, and short-wave infrared to generate a spectral feature matrix, a pre-set feature selection algorithm is used to construct a set of crop classification models, enabling the model to fully utilize multi-band spectral information for classification. Furthermore, the training process of the feature fusion association model utilizes a dynamic learning rate adjustment mechanism and a gradient descent algorithm for hyperparameter optimization. This allows the model to automatically adjust model parameters based on historical data, improving its generalization and adaptability. This dynamic optimization mechanism enables the model to better adapt to the characteristics of satellite imagery data from different regions and time phases, maintaining high classification accuracy even in complex environments.

[0047] During feature optimization and processing, real-time spectral and texture sets are input, combined with the classification accuracy target, to enhance features. An evolutionary algorithm is then used to create a feature optimization population, which is then subjected to multiple rounds of evolution to ultimately extract the global optimal solution as the optimization result. This optimization algorithm-based feature enhancement technology automatically adjusts feature parameters based on the characteristics of the current image data, further improving the feature's discriminative capabilities and ensuring that the classification results meet accuracy requirements. Compared to traditional fixed-parameter feature processing methods, the present invention's dynamic feature optimization mechanism can more flexibly adapt to varying data conditions, improving classification reliability.

[0048] In terms of geometric calibration and spatial relationship processing, a coordinate affine transformation algorithm is used to calibrate the spatial relationships within the target classification scheme. This algorithm dynamically adjusts feature matching parameters based on satellite positioning accuracy and image distortion coefficients, effectively correcting the geometric distortion and radiometric errors of satellite imagery, improving the accuracy of feature extraction and the spatial positioning accuracy of classification results. By precisely calibrating the spatial coordinates of the feature matching sequence, the accuracy and reliability of the classification results in the spatial dimension are ensured, providing strong support for the precise demarcation and dynamic monitoring of crop planting areas.

[0049] Furthermore, the present invention introduces temporal feature association rules. By acquiring crop growth cycle data from images of different time phases, temporal feature association rules are established and used to optimize the spatial distribution of crop types in the classification scheme. This innovation enables the classification method to fully utilize the temporal dimension of crop growth, dynamically reflecting changes in crop growth, improving the continuity and reliability of classification results over time, and providing more comprehensive information support for dynamic management and decision-making in agricultural production. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a working principle diagram of the crop classification method based on multi-source satellite images according to the present invention;

[0051] Figure 2 Design diagram for feature fusion correlation model training;

[0052] Figure 3 Design diagram for iterative optimization of the model;

[0053] Figure 4 Design diagram of the feature enhancement evolutionary algorithm. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] See also Figures 1-4 The present invention relates to a crop classification method based on multi-source satellite images, and the specific implementation steps are as follows:

[0056] S1: Acquire multi-temporal satellite imagery and construct a classification model ensemble: Acquire multi-temporal satellite imagery covering the target area, extract spectral reflectance characteristics of different bands, and form a spectral signature matrix containing spectral signature parameters. Based on this spectral signature matrix, a specific algorithm is used to construct a crop classification model ensemble, providing a model foundation for subsequent classification.

[0057] S2: Set Feature Parameter Categories and Build a Feature Fusion Association Model: We set feature parameter categories corresponding to various land cover types to form a feature parameter category set. Based on this category set, we extract texture feature parameters, vegetation index parameters, and spatial resolution data from multiple sets of historical image data. Through data association and model training, we build a feature fusion association model capable of integrating multi-dimensional features.

[0058] S3: Extract and optimize current image features: Based on the feature parameter category set, extract the spectral feature parameters and texture feature parameters of each crop area from the current satellite image, and optimize the features by combining them with the established feature fusion association model to generate optimized spectral feature sets and texture feature sets, thereby improving the classification effectiveness of feature data.

[0059] S4: Update classification logic and calibrate spatial relationships: Based on the optimized spectral feature set, texture feature set, and crop classification model set, the classification logic of the current image is updated to generate a target classification scheme. A specific algorithm is used to calibrate the spatial relationships within the feature extraction process in the target classification scheme to ensure the spatial accuracy of the classification results.

[0060] The present invention will be further described below in conjunction with Examples 1 to 5:

[0061] Example 1: This example describes step S1 in detail. S1 includes the process of defining a set of crop types, collecting specific band data to generate a spectral feature matrix, and constructing a set of crop classification models using a feature selection algorithm, as follows:

[0062] Execute step S11: Define multiple crop types and form a crop type set. In actual operation, the specific crop types need to be determined based on the agricultural planting structure of the target area. For example, if the target area is the main grain-producing area of the North China Plain, the crop types may include typical crops such as winter wheat, summer corn, and cotton; if it is the southern hilly area, rice, rapeseed, tea, etc. can be included. The definition of the crop type set must cover all major planting types in the target area to ensure the comprehensiveness of the classification objects. After the definition is completed, a unique identification category must be established for each crop to facilitate subsequent data association and model training.

[0063] Data is collected in the corresponding near-infrared, red-edge, and short-wave infrared bands according to the crop type. Different satellite imagery bands carry different ground feature information. The near-infrared band is sensitive to vegetation chlorophyll content and leaf structure. The red-edge band (typically in the 680-760nm range) represents the transition from visible light to near-infrared spectral reflectance, reflecting vegetation growth and species differences. Short-wave infrared bands (such as 1400-1900nm and 2000-2500nm) are associated with vegetation moisture content and biochemical composition. For each crop type, satellite imagery data covering the target area must be collected at multiple phases throughout its growth cycle (e.g., seedling, flowering, and maturity) to ensure spectral characteristics are captured at different growth stages.

[0064] During data collection, attention must be paid to satellite sensor selection and data preprocessing. For example, multi-source satellite data, such as the Landsat series (equipped with an OLI sensor covering near-infrared and shortwave infrared bands) and Sentinel-2 (providing 13 bands, including a red-edge band), can be used to meet multi-band acquisition requirements. The collected raw image data requires preprocessing operations such as radiometric calibration, atmospheric correction, and geometric correction to eliminate the effects of sensor noise, atmospheric scattering, and terrain undulation on spectral reflectance, ensuring the accuracy and comparability of the extracted spectral characteristic parameters.

[0065] After preprocessing, the spectral reflectance values of each pixel in the near-infrared, red-edge, and shortwave infrared bands were extracted for each phase of each crop type image. Taking the pixel as the basic unit, the reflectance values of the three bands were combined into a feature vector, which contains the spectral characteristic information of the ground feature corresponding to that pixel. For example, for a pixel with a reflectance of 0.45 in the near-infrared band, 0.38 in the red-edge band, and 0.22 in the shortwave infrared band, the feature vector [0.45, 0.38, 0.22] was formed. The feature vectors of all pixels belonging to the same crop type were aggregated and combined with their corresponding crop category labels to construct a spectral feature subset for that crop type.

[0066] The spectral feature subsets for all crop types are integrated to form a global spectral feature matrix. The rows of this matrix represent different pixel samples, and the columns represent the reflectance characteristics of the near-infrared, red-edge, and shortwave infrared bands. Each row of samples corresponds to a specific crop category. The structure of the spectral feature matrix must meet the input requirements of the machine learning model and is typically stored in a two-dimensional table, where the sample label column clearly identifies the crop type to which each pixel belongs.

[0067] After the spectral feature matrix is constructed, step S12 is entered: the spectral feature matrix is input into the classification model generation module, where a set of crop classification models is constructed using a pre-set feature selection algorithm. The classification model generation module can be an integrated learning platform or a customized algorithm framework. Its core function is to receive the spectral feature matrix, select the most discriminative feature combinations for classification using a feature selection algorithm, and then construct multiple models based on different classification algorithms.

[0068] The purpose of the pre-set feature selection algorithm is to select features that contribute most to crop classification from the spectral features of the three bands, thereby reducing the dimensionality of the feature space, avoiding the "curse of dimensionality," and improving model efficiency. Common feature selection algorithms include recursive feature elimination (RFE), random forest feature importance ranking, and variance selection. Taking the recursive feature elimination algorithm as an example, its basic process is as follows: first, an initial classification model (such as a support vector machine (SVM)) is trained using features from all three bands. The importance score of each feature is calculated, and then the feature with the lowest score is removed. The model is retrained on the remaining features, and this process is repeated until the preset number of features is reached or a stopping condition is met. This process can determine the optimal feature combination; for example, a combination of the near-infrared band and the red-edge band may be selected as the key feature.

[0069] Based on feature selection, we use different classification algorithms to construct multiple crop classification models. Optional classification algorithms include support vector machines (SVMs), random forests (RandomForests), gradient boosted decision trees (GBDTs), and neural networks (such as multi-layer perceptrons (MLPs). Each algorithm has a different hypothesis space and learning capabilities. By building a model ensemble, we can achieve complementary advantages and improve classification robustness and accuracy.

[0070] Taking the random forest model as an example, its construction process involves randomly extracting multiple sample subsets (bootstrap sampling) from the spectral feature matrix. Within each sample subset, a subset of features (such as the near-infrared and red-edge bands after feature selection) is randomly selected to construct a decision tree. The final classification result is determined through a voting mechanism among these multiple decision trees. The random forest model is robust to noisy data and overfitting, making it suitable for processing high-dimensional feature data.

[0071] For the support vector machine model, the spectral feature matrix must first be standardized, and each eigenvalue must be scaled to the same range (such as [0,1] or [-1,1]). Then, the features are mapped to a high-dimensional space through a kernel function (such as the radial basis kernel function RBF) to find the hyperplane that maximizes the classification interval to achieve classification of different crop types.

[0072] When building a collection of crop classification models, each model needs to be trained, validated, and tuned. Specifically, the spectral feature matrix is divided into a training set and a validation set (e.g., a 7:3 ratio). The training set is used to learn the model parameters. The validation set is used to evaluate the model's classification accuracy, recall rate, F1 value, and other indicators. Model hyperparameters (such as the number of trees in the random forest, the penalty coefficient C of the SVM, and the kernel function parameter γ) are adjusted until the model performance on the validation set reaches a stable state.

[0073] The above process ultimately results in a collection of crop classification models constructed from multiple different algorithms. Each model corresponds to a different feature combination or algorithm parameter, enabling modeling of crop spectral characteristics from different perspectives. The advantage of this model collection is that it can combine the prediction results of multiple models through ensemble learning methods (such as voting and stacking), further improving classification performance.

[0074] Example 2: This example describes step S2 in detail. S2 includes setting a feature parameter category set, obtaining multi-dimensional feature parameters from historical image data, and establishing a feature fusion association model through model training, as follows:

[0075] Execute step S21: set multiple feature parameter categories and form a feature parameter category set. The feature parameter categories must cover key indicators that can reflect the differences in surface cover types, including normalized vegetation index thresholds, spatial texture complexity coefficients, and surface temperature ranges. Normalized vegetation indices (such as NDVI and EVI) can quantify vegetation coverage and growth conditions, spatial texture complexity coefficients are used to describe the uniformity or roughness of surface texture (such as contrast and entropy calculated by gray-level co-occurrence matrix), and surface temperature ranges are related to vegetation transpiration and energy balance. The selection of these parameter categories must be based on prior knowledge of crop spectral and texture characteristics to ensure their effectiveness in crop classification.

[0076] After forming a set of characteristic parameter categories, texture feature parameters, vegetation index parameters, and spatial resolution data are obtained from multiple sets of historical imagery data based on the crop type and characteristic parameter category sets. This historical imagery data must cover multiple years and seasons for the target area, including images of different crop types at different growth stages. For example, for areas rotating between winter wheat and summer corn, imagery data must be collected at multiple time periods, including the sowing period, tillering period, jointing period, and grain filling period, to ensure that characteristic changes throughout the crop's growth cycle are covered.

[0077] To extract texture feature parameters, the historical imagery is first preprocessed (e.g., noise removal and image enhancement). Texture analysis algorithms (e.g., gray-level co-occurrence matrix (GLCM) and local binary pattern (LBP)) are then used to calculate the texture features of each pixel or region. Taking the gray-level co-occurrence matrix as an example, by statistically analyzing the frequency of gray-level co-occurrences of pixels at specific directions and distances within the image, parameters such as contrast (reflecting texture clarity), entropy (reflecting texture complexity), and angular second moment (reflecting texture uniformity) can be obtained. For each phase of imagery for each crop type, texture feature parameters are extracted and a texture feature vector is generated.

[0078] Vegetation index parameters are extracted based on multispectral data from historical imagery. Taking NDVI as an example, its calculation formula is (near-infrared reflectance - red reflectance) / (near-infrared reflectance + red reflectance), which reflects the chlorophyll content and photosynthesis intensity of vegetation. For each crop type, vegetation indices such as NDVI and EVI are calculated for imagery at different temporal phases to form a composite vegetation index data set. Furthermore, spatial resolution data must record the ground pixel resolution of the corresponding image (e.g., 10 meters, 20 meters) to facilitate scale matching during subsequent feature fusion.

[0079] In parallel with acquiring the aforementioned data, corresponding crop identification accuracy data must be collected. This data can be obtained through ground-based surveys, agricultural statistical data, or manual interpretation. It is used to label the actual crop types in each region of historical imagery and serve as ground truth labels for model training. For example, sample point coordinates can be obtained through field sampling, and the corresponding crop category can be determined based on image data to establish a mapping between feature parameters and crop types.

[0080] Based on the above data, a historical feature matrix, a historical vegetation matrix, and a historical accuracy dataset were constructed. The rows of the historical feature matrix represent samples (such as plots or pixels in an image), and the columns correspond to texture feature parameters (such as contrast and entropy). Each row records the texture feature vector of a sample and the corresponding crop category label. The historical vegetation matrix stores the vegetation index combination data (such as NDVI and EVI values) and category labels for each sample. The historical accuracy dataset contains the classification accuracy reference value of each sample under a specific feature combination. This value can be calculated by comparing the model prediction results with the true value labels (such as a comprehensive evaluation value of indicators such as precision and recall).

[0081] After completing the data construction, enter step S22: Use the historical feature matrix, historical vegetation matrix, and historical accuracy data set to train the initial feature fusion association model to obtain the feature fusion association model. This process uses machine learning algorithms to achieve multi-dimensional feature fusion and classification accuracy prediction. The specific steps are as follows:

[0082] Execute step S221: Establish an initial feature fusion association model based on the set of crop types and the set of feature parameter categories. The initial model can adopt architectures such as support vector machines (SVMs), neural networks (such as multi-layer perceptrons (MLPs), or ensemble learning models (such as gradient boosting machines (GBMs)). Taking a neural network as an example, the number of nodes in its input layer corresponds to the total dimension of the texture feature parameters and vegetation index parameters (for example, if the texture features contain 4 parameters and the vegetation index contains 2 parameters, the input layer will have 6 nodes). The hidden layer uses an activation function (such as ReLU) to achieve nonlinear feature transformation. The number of nodes in the output layer corresponds to the number of crop types, and the classification probability is output through the Softmax function.

[0083] Execute step S222: Input the historical feature matrix and the historical vegetation matrix row by row into the initial feature fusion association model to generate a historical accuracy prediction dataset. Before inputting the data, the texture features and vegetation index must be normalized (e.g., to the range [0, 1]) to prevent the impact of different feature scales on model training. The model uses forward propagation to fuse the texture and vegetation index features of each sample and outputs a classification accuracy prediction value for that sample. This value reflects the model's estimate of the classification accuracy given that sample's feature combination.

[0084] Execute step S223: Set the precision error threshold and calculate the cross-entropy error between the historical precision prediction dataset and the historical precision dataset. The precision error threshold is a measure of whether the model's prediction accuracy meets the target and should be set based on the actual requirements of the classification task (e.g., 0.1 or 0.2). The cross-entropy error quantifies the difference between the predicted value and the true value. It is calculated by summing the logarithmic loss of the true precision value and the predicted precision value for all samples and taking the average.

[0085] When the cross-entropy error exceeds the precision error threshold, the model parameter correction mechanism is triggered, and the gradient descent algorithm is used to iteratively correct the hyperparameters of the initial feature fusion association model. The gradient descent algorithm calculates the gradient of the error with respect to the model parameters and updates the parameters in the opposite direction of the gradient to minimize the error. During the iterative process, the following details should be noted:

[0086] Batch processing: Stochastic gradient descent (SGD) or mini-batch gradient descent (Mini-Batch GD) can be used to divide the training data into multiple batches and gradually update the parameters to avoid excessive memory usage and low computational efficiency.

[0087] Learning rate control: The learning rate determines the step size of parameter updates. The initial learning rate can be set to a small value, such as 0.01, to ensure the stability of parameter updates. As iterations proceed, the learning rate can be gradually reduced using a learning rate decay strategy (for example, multiplying the learning rate by 0.9 after every 50 iterations) to avoid late-stage parameter fluctuations.

[0088] Regularization: To prevent model overfitting, L1 or L2 regularization terms can be added to the loss function to penalize the complexity of model parameters and improve the generalization ability of the model.

[0089] During the iteration process, the trend of the cross-entropy error is regularly evaluated. When the error decreases by less than a preset threshold (e.g., 0.001) over multiple consecutive iterations or reaches the maximum number of iterations (e.g., 1000), the model is considered converged and iterations are stopped, ultimately forming a feature fusion association model. This model effectively integrates texture features, vegetation index features, and spatial resolution information, establishing a mapping relationship between multi-dimensional features and crop classification accuracy, providing support for subsequent feature optimization of the current image.

[0090] During model training, it's also important to ensure that data is properly partitioned. Historical data is typically divided into a training set (approximately 70%), a validation set (approximately 20%), and a test set (approximately 10%). The training set is used for learning model parameters, the validation set is used for adjusting hyperparameters and monitoring overfitting, and the test set is used to evaluate the model's generalization capabilities. Cross-validation (such as k-fold cross-validation) can further enhance the reliability of model evaluation and ensure that the feature fusion correlation model exhibits stable performance across different data subsets.

[0091] Example 3: This example describes in detail the hyperparameter correction process in step S223, focusing on the application of the dynamic learning rate adjustment mechanism in the gradient descent algorithm.

[0092] During the training of a feature fusion correlation model, when the cross-entropy error exceeds the precision error threshold, the model's hyperparameters (such as the neural network's weights, biases, or regularization coefficients) must be iteratively corrected using a gradient descent algorithm. Traditional gradient descent algorithms with fixed learning rates can suffer from slow initial convergence or later parameter fluctuations. Therefore, a dynamic learning rate adjustment mechanism is introduced. Its core idea is to automatically adjust the learning rate based on the training progress to balance convergence speed and optimization accuracy.

[0093] The specific implementation of the dynamic learning rate adjustment mechanism is as follows: First, define the functional relationship between the learning rate and the number of iterations. Assume that the current number of iterations is , the initial learning rate is , the maximum number of iterations is , then the learning rate It can be expressed as:

[0094]

[0095] in, is the learning rate decay function, which describes how the learning rate changes with the number of iterations. Common decay functions include exponential decay, polynomial decay, and cosine decay. The following describes their implementation details.

[0096] Exponential decay: The expression of the exponential decay function is:

[0097]

[0098] in, is the decay rate ( ), needs to be pre-set according to the model training situation. For example, if , the learning rate decays to 95% of the previous one at each iteration. This method has the advantages of being simple to implement and can quickly reduce the learning rate to avoid late oscillations, but the decay rate is fixed and may not be suitable for complex training curves.

[0099] In practical applications, exponential decay can be combined with a warm-up mechanism. Iterations), the learning rate does not decay immediately, but increases linearly to , to help the model quickly explore the parameter space in the early stage. The learning rate expression in the warm-up phase is:

[0100]

[0101] After the warm-up is complete, the model enters the exponential decay phase. This approach is suitable for scenarios where model parameters are poorly initialized and can improve initial convergence stability.

[0102] Polynomial decay: The polynomial decay function describes the decreasing trend of the learning rate through a polynomial function. Its general form is:

[0103]

[0104] in, is the degree of the polynomial (a positive integer). When , it degenerates into linear attenuation; when When , the learning rate decreases rapidly in the early stage and tends to be flat in the later stage. For example, if , then the learning rate is in the middle of the iteration ( ), which decays to 25% of the initial value. It is suitable for models that require fast initial convergence and fine-tuning in the later stage.

[0105] The advantage of polynomial decay is that it can be adjusted The value flexibly controls the decay speed, but the maximum number of iterations needs to be determined in advance If the training converges prematurely, computing resources may be wasted. Therefore, an early stopping mechanism can be used to terminate the training early when the validation set error stops decreasing, thus avoiding invalid iterations.

[0106] Cosine decay: The cosine decay function simulates the downward trend of the cosine curve. Its expression is:

[0107]

[0108] This function decreases slowly in the early stages, accelerates in the middle stages, slows down again in the later stages, and ultimately approaches zero. A characteristic of cosine decay is that the learning rate fluctuates slightly around its minimum value in the later stages of training. This may help the model escape from local optima and improve generalization. For example, smaller learning rate fluctuations in the later stages of training allow for more precise parameter adjustments near the optimal solution.

[0109] To further enhance the flexibility of cosine decay, periodic cosine annealing can be introduced. This means that after the learning rate drops to a certain level, it returns to a higher value and decays again, forming multiple decay cycles. The maximum number of iterations within each cycle can be reduced by a certain percentage (for example, the length of each cycle is halved) to simulate the annealing process, helping the model search for optimal solutions at different scales.

[0110] In the implementation of the dynamic learning rate adjustment mechanism, the following key links need to be noted:

[0111] Learning rate monitoring: After each iteration, record the current learning rate and the corresponding cross entropy error, draw a learning rate-error curve, and observe the correlation between the two. If the error continues to rise after the learning rate decreases, it may indicate that the learning rate decays too quickly, and the parameters of the decay function need to be adjusted (such as increasing or reduce ).

[0112] Synchronous parameter updates: In deep learning frameworks such as TensorFlow and PyTorch, dynamic learning rate adjustments are typically implemented through optimizers. For example, PyTorch's torch.optim.lr_scheduler module provides a variety of learning rate schedulers (such as ExponentialLR and CosineAnnealingLR), which automatically update the learning rate based on the number of iterations and ensure synchronized adjustments of model parameters and learning rates.

[0113] Hyperparameter Co-Optimization: Parameters in the dynamic learning rate adjustment mechanism (such as 、 、 ) needs to be optimized in conjunction with other model hyperparameters (such as the number of hidden layer neurons and the regularization coefficient). GridSearch or RandomSearch can be used to find the optimal parameter combination on the validation set.

[0114] Taking the neural network model as an example, assuming the initial learning rate , using exponential decay function, decay rate , maximum number of iterations In the early stages of training (the first 100 iterations), the learning rate gradually decays from 0.01 to about ; In the middle of training (500th iteration), the learning rate is reduced to about In the later stages of training, the learning rate approaches 0. As the learning rate decreases, the step size of gradient descent gradually decreases, and the parameter update shifts from rapid global search to fine local adjustment, ensuring that the model eventually converges to a better parameter space.

[0115] The introduction of a dynamic learning rate adjustment mechanism makes the hyperparameter modification process more adaptive and intelligent. Compared with a fixed learning rate, its advantages are:

[0116] Fast initial convergence: A higher learning rate can accelerate the update of parameters in the initial stage and shorten the training time.

[0117] Late-stage stable optimization: The attenuated learning rate can prevent the parameters from oscillating back and forth near the optimal solution, thereby improving convergence accuracy.

[0118] Enhanced robustness: Through adaptive adjustment, it can cope with the differences in feature distribution of different data sets and reduce the cost of manual parameter adjustment.

[0119] In specific implementations, it is necessary to select an appropriate decay strategy based on factors such as the data size of the historical feature matrix and the historical vegetation matrix, model complexity, etc. For example, for small datasets or simple models, linear decay or exponential decay may be sufficient; for large datasets or deep neural networks, cosine decay or periodic decay may be more effective.

[0120] Example 4: This example elaborates on step S3 and describes the entire process of extracting features from the current satellite image, optimizing features, and generating an optimized feature set in conjunction with a specific application scenario. The details are as follows:

[0121] Suppose the target area is an agricultural demonstration zone in a certain province, and crop classification is currently required using Sentinel-2 satellite imagery from May 2025. First, execute step S31: obtain the actual spectral and texture characteristic parameters of each plot area in the current satellite imagery, generating a real-time spectral set and texture set. In practice, image data is loaded using remote sensing image processing software (such as ENVI or QGIS), and the image is segmented into different plot units using vector boundary files (such as plot distribution maps provided by agricultural departments). For example, for plot number A101, its reflectance values in the blue, green, red, and near-infrared bands are extracted as spectral characteristic parameters. Simultaneously, a gray-level co-occurrence matrix algorithm is used to calculate the plot's texture characteristic parameters, including contrast, entropy, and homogeneity, reflecting the planting uniformity or foliar structural differences of crops within the plot. Repeat the above operation for all plots to form a real-time spectrum set (recording the reflectance of each band) and a real-time texture set (recording texture feature parameters) containing thousands of plot samples. Each sample is associated with its plot number and geographic location coordinates.

[0122] Next, proceed to step S32: Set the classification accuracy target and initiate the feature optimization process. Assuming the classification accuracy target is set to 90% based on business requirements, the real-time spectrum set and real-time texture set are input into a trained feature fusion association model (such as the neural network model established in Example 2). The model calculates the current classification accuracy prediction value through forward propagation. If the prediction value is 85% (which does not reach the target value), the feature enhancement process is triggered, which is implemented through the following steps:

[0123] S321: Create a feature optimization population and initialize parameters. Define the feature optimization population as a set of potential feature adjustment schemes, with each individual representing a combination of adjustments to spectral and texture feature parameters. For example, the population size is set to 50 individuals, the maximum number of evolutionary generations is 100, and the current number of evolutionary generations is initialized to 0. The genes of each individual encode the adjustment range of the corresponding spectral feature parameter (such as near-infrared reflectance) and texture feature parameter (such as contrast). For example, using real number encoding, the value of each gene represents the adjustment ratio of the corresponding feature (such as +5%, -3%, etc.).

[0124] S322: Set feature adjustment ranges and generate an initial population position matrix. Based on the physical significance of spectral and texture features and historical data distribution, set the adjustment ranges for each feature parameter. For example, the adjustment range for near-infrared reflectance is ±10% (to avoid exceeding the normal reflectance range of vegetation), and the adjustment range for texture contrast is ±15% (to account for texture variations during crop growth). The initial population position matrix is generated using uniform random sampling, with each row vector corresponding to a feature adjustment plan for an individual. For example, the adjustment plan for the first individual might be: near-infrared reflectance +8%, red-edge reflectance -5%, texture entropy +10%. The adjustment plan for the second individual might be: near-infrared reflectance -3%, red-edge reflectance +6%, texture contrast -8%, and so on. This ensures that the initial population covers different regions of the feature space.

[0125] S323: Construct a population fitness function. This function aims to improve classification accuracy, using the prediction accuracy of the feature fusion association model for the adjusted features as the fitness value. Specifically, for each individual's feature adjustment scheme, the real-time spectral set and real-time texture set are adjusted according to the scheme (for example, multiplying the near-infrared reflectance value by 1.08 and adding 10% of the original value to the texture entropy value). This generates an adjusted feature set, which is then input into the model to calculate a new classification accuracy prediction value, which serves as the individual's fitness value. For example, if the adjusted prediction accuracy of one individual is 87%, its fitness value is 0.87; if the adjusted prediction accuracy of another individual is 89%, its fitness value is 0.89. The latter individual has a higher survival competitiveness in the population.

[0126] S324: Multiple rounds of evolution update the population position, using the basic operations of genetic algorithms (selection, crossover, mutation) to evolve the population:

[0127] Selection: Rank individuals based on their fitness, using roulette wheel selection or tournament selection to retain individuals with higher fitness (e.g., the top 30 individuals) for the next generation and eliminate individuals with lower fitness. For example, tournament selection randomly selects five individuals each time, retains the one with the highest fitness, and repeats until the next generation population size is reached.

[0128] Crossover: Selected individuals are crossovered pairwise, exchanging some of their genetic fragments to generate new individuals. For example, for a feature adjustment scheme involving two individuals, a single-point crossover is performed at the third gene position (e.g., texture entropy adjustment), generating two offspring individuals whose genes inherit the first and second halves of the parent's adjustment parameters, respectively. The crossover probability is typically set to 0.8-0.9 to ensure population diversity.

[0129] Mutation: Genetic mutation is performed on individuals after crossover with a low probability (e.g., 0.01-0.05), randomly changing the adjustment range of a certain feature to prevent the population from prematurely converging to a local optimum. For example, the adjustment range of the red-edge reflectance of an individual is +6%, which may change to +9% or -2% after mutation, thereby exploring new areas of the feature space.

[0130] After each evolutionary cycle, the fitness values of all individuals are recalculated, and the global optimal solution (i.e., the feature adjustment solution with the highest current fitness) is recorded. For example, after the 10th generation of evolution, the prediction accuracy corresponding to the global optimal solution increased to 88%; by the 30th generation, it reached 89.5%, close to the target value.

[0131] S325: Terminate the evolution process and extract the optimization results. The evolution process terminates when the number of evolutionary generations reaches the maximum (e.g., generation 100) or when the fitness value changes by less than 0.1% over multiple generations (e.g., five generations). A feature adjustment scheme corresponding to the global optimal solution is extracted and applied to the real-time spectral and texture sets to generate the optimized spectral and texture feature sets. For example, the global optimal solution might be: near-infrared reflectance +7%, red-edge reflectance -4%, texture contrast -5%, and texture entropy +12%. After applying this adjustment scheme, the model's prediction accuracy reaches 90.2%, meeting the target requirements.

[0132] During feature enhancement, the following key details should be noted:

[0133] Physical constraints: When adjusting spectral feature parameters, ensure that the adjusted values are within the reasonable measurement range of the satellite sensor. For example, reflectance values must be between 0 and 1, avoiding negative values or values exceeding 1. Adjustments to texture feature parameters must conform to the actual texture characteristics of the ground object (for example, the texture complexity of crop leaves should not change dramatically with adjustments).

[0134] Computational efficiency optimization: When the image data volume is large (e.g., containing hundreds of thousands of plots), a distributed computing framework (such as Spark) can be used to parallelize the fitness calculation of individual populations, reducing training time. In addition, an early stopping mechanism can be set to terminate the evolution early if the fitness value reaches the target value mid-evolution, saving computing resources.

[0135] Collaborative adjustment of multiple features: Spectral and textural features are inherently correlated (for example, densely vegetated areas typically have higher near-infrared reflectivity and lower texture complexity). Therefore, when adjusting features, it is important to avoid over-adjusting a single feature, which could distort the physical meaning. For example, if near-infrared reflectivity increases significantly, it is necessary to simultaneously consider whether texture complexity conforms to vegetation growth patterns to avoid generating an unreasonable feature combination of "high reflectivity + high texture complexity."

[0136] To illustrate the practical application of feature optimization, consider a field of corn and soybeans. Suppose the original spectral signature of a field has a near-infrared reflectance of 0.4, a red-edge reflectance of 0.35, and a texture contrast of 150 (normalized values calculated based on the gray-level co-occurrence matrix). The model initially predicts "soybean," but actual ground surveys indicate "corn." Through evolution of the feature-optimized population, it may be found that adjusting the near-infrared reflectance to 0.45 (a 12.5% increase) and the texture contrast to 130 (a 13.3% decrease) corrects the model's prediction to "corn," indicating that this adjustment effectively enhances the distinguishability between corn and soybeans.

[0137] In another scenario, the original texture entropy value of a wheat field was 80 (reflecting leaf structural complexity), and the NDVI value of the spectral signature was 0.65. The model's initial prediction accuracy was low. After feature optimization, the NDVI value was adjusted to 0.68 (+4.6%) and the texture entropy value to 75 (-6.25%), correspondingly improving prediction accuracy from 82% to 91%. This demonstrates that the coordinated adjustment of spectral and textural features effectively highlights the characteristic differences of wheat during the jointing stage.

[0138] Example 5: This example elaborates on step S4 and the extended step S5 in detail, and illustrates the entire process of generating a target classification scheme, calibrating spatial relationships, and optimizing temporal features in conjunction with specific application scenarios, as follows:

[0139] Assume that the target area is an agricultural area in a certain watershed, containing three major crops: wheat, corn, and rice. Currently, the Sentinel-2 satellite imagery from May 2025 (spatial resolution 10 meters) and the Landsat-8 satellite imagery from June 2025 (spatial resolution 30 meters) need to be processed. First, step S41 is executed: the optimized spectral feature set and texture feature set are loaded into the image classification unit to generate a target classification scheme. The image classification unit integrates the crop classification model set constructed in Example 1 (such as the random forest model and the support vector machine model), uses the optimized features as input, and performs classification prediction for each plot.

[0140] For example, plot B203 was optimized for spectral features including a near-infrared reflectance of 0.48 and a red-edge reflectance of 0.39. Its texture features included a contrast ratio of 120 and an entropy of 75. The random forest model, using multiple decision trees, voted and predicted the plot as "wheat." The support vector machine model, using radial basis kernel function calculations, also predicted the same. Combining the outputs of the ensemble of models, the target classification scheme labeled the plot "wheat." This process was repeated for all plots in the region, generating preliminary classification results that included crop types for each plot, such as "Plot B101: corn, Plot B102: rice, Plot B103: wheat."

[0141] Next, step S42 is executed: the spatial coordinates of the feature matching sequence in the target classification scheme are identified and geometrically calibrated according to the preset spatial resolution constraint. The spatial coordinates of the feature matching sequence are derived from the geographic positioning information of the satellite image. For example, the coordinates of the upper left corner of plot B203 are (116.5°E, 39.2°N), and the coordinates of the lower right corner are (116.501°E, 39.199°N). The preset spatial resolution constraint is set based on the actual resolution of the satellite image. For example, the 10-meter resolution of the Sentinel-2 image requires that the boundary error between adjacent plots in the classification result does not exceed half a pixel (5 meters).

[0142] The coordinate affine transformation algorithm is used for geometric calibration, which is implemented by the following steps:

[0143] Ground control point selection: Select evenly distributed ground control points (GCPs) in the study area, such as road intersections, farmland boundary turning points, etc., and obtain their real coordinates through GPS measurement ( ), and mark the corresponding pixel coordinates on the image ( For example, select 10 control points, covering the four corners and the center of the area.

[0144] Transformation parameter calculation: According to the image coordinates and real coordinates of the control points, establish the affine transformation equations and solve the translation ( ), Rotation ( ), Zoom ( ) and other parameters to minimize the deviation between the image coordinates and the real coordinates.

[0145] Coordinate correction: Substitute the spatial coordinates of all plots in the target classification scheme into the affine transformation formula to generate corrected coordinates. For example, the original coordinates of the upper left corner of plot B203 are adjusted to (116.5005° East, 39.2002° North) after correction, which is closer to the actual location.

[0146] The coordinate affine transformation algorithm dynamically adjusts feature matching parameters based on satellite positioning accuracy (e.g., Sentinel-2 positioning accuracy is 5 meters) and image distortion coefficients (obtained through sensor parameters). For example, if the image distortion coefficient is large due to uneven terrain in a certain area, the number of control points can be increased and the calibration window can be narrowed to improve local correction accuracy. If satellite positioning accuracy is high, the density of control points can be appropriately reduced to improve processing efficiency.

[0147] After geometric calibration, the expansion step S5 begins: obtaining crop growth cycle data for images at different time phases within the target classification scheme, establishing temporal feature association rules, and optimizing the spatial distribution map of crop types. Taking winter wheat as an example, its growth cycle includes sowing (October), tillering (November), greening (March), jointing (April), heading (May), and maturity (June). Spectral characteristics (such as NDVI) and textural features (such as leaf surface complexity) of images at different time phases exhibit regular changes with growth stage: During the sowing phase, the ground is covered with soil, resulting in lower NDVI values (approximately 0.2) and higher texture complexity (reflecting the rough surface of plowed soil); during the tillering phase, vegetation cover increases, with NDVI values rising to 0.4-0.5 and texture complexity decreasing (leaf growth begins to even out); during the jointing phase, NDVI values reach 0.7-0.8, and texture contrast increases (increasing structural differences between stems and leaves).

[0148] Based on these rules, temporal feature association rules were established: if a plot of land was classified as "wheat" in May imagery, its NDVI value in June imagery should show a downward trend (entering maturity, with leaves beginning to wither and turn yellow), and its texture contrast should further increase (the ear structure increases texture complexity). Conversely, if a plot of land was classified as "corn" in May (the planting season is usually in June), its May imagery should show low NDVI values (bare ground or covered with weeds) and high texture complexity (unplanted, rough ground or overgrown with weeds).

[0149] The specific process of applying temporal feature association rules to optimize the spatial distribution map is as follows:

[0150] Multi-temporal data matching: Spatial registration is performed on the current classification result (May 2025) with the previous temporal image (e.g. November 2024) and the next temporal image (June 2025) to ensure that the coordinates of the same plot in different temporal images are consistent.

[0151] Time Series Analysis: For each plot, spectral features (such as reflectance and NDVI in each band) and texture features (such as contrast and entropy) at different time phases are extracted to generate a time series characteristic curve. For example, the NDVI for plot B301 was 0.3 in November 2024, 0.85 in May 2025, and 0.78 in June 2025. This corresponds to the NDVI variation pattern of winter wheat from tillering to maturity, and the classification result of "wheat" is reasonable.

[0152] Abnormal classification correction: If a plot of land is classified as "rice" in May 2025, but its NDVI value in November 2024 is 0.7 (significantly higher than the surface characteristics of rice during the winter fallow period), the anomaly detection mechanism will be triggered. Combined with historical planting pattern data (rice in this area is usually sown in spring), it will be judged that the classification result may be wrong. The classification model will be re-called for review and corrected to "wheat" or other reasonable types.

[0153] Spatial distribution optimization: Based on the consistency constraints of temporal features, the classification results of adjacent plots are adjusted. For example, if the classification results for a certain area in May 2025 show interlaced "wheat" and "corn" plots, but historical temporal data indicates that the area is a continuous wheat cropping area with no corn planting records, the corn plots will be corrected to wheat to ensure that the spatial distribution map is consistent with the actual planting pattern.

[0154] For example, consider optimizing the classification of rice and corn. Suppose that plot C101 was classified as "corn" in May imagery. However, its NDVI value in June 2025 imagery was 0.9 (significantly above the normal range for corn seedlings) and had low texture contrast (uniform leaf growth, consistent with the tillering stage of rice). Analysis of temporal feature association rules revealed that the plot was likely in the rice transplanting phase (low NDVI) in May and entering the tillering phase (increased NDVI and uniform texture) in June. Therefore, the classification result was revised to "rice," avoiding errors caused by misclassification based on single temporal features.

[0155] In another scenario, plot C202 was classified as "fallow" in May 2025, but its NDVI value in November 2024 was 0.6 (approaching the wheat tillering stage), and in June 2025 it was 0.85 (approaching wheat maturity). Through temporal feature continuity analysis, it was determined that wheat was actually planted on this plot. The May classification result was incorrectly classified due to sensor noise or insufficient feature optimization. After correction, it was included in the wheat classification area, improving the accuracy of the spatial distribution map.

[0156] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0157] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A crop classification method based on multi-source satellite images, characterized in that: The following steps are involved: S1. Acquire multi-temporal satellite image data, extract spectral reflectance characteristics of different bands, and generate a spectral feature matrix; Construct a set of crop classification models based on the spectral feature matrix; S2. setting characteristic parameter categories corresponding to a plurality of land cover types to form a characteristic parameter category set; Based on the feature parameter category set, texture feature parameters, vegetation index parameters and spatial resolution data from multiple sets of historical image data are obtained, and a feature fusion association model is established; S3, extracting spectral characteristic parameters and texture characteristic parameters of each crop area in the current satellite image according to the feature parameter category set, performing feature optimization in combination with the feature fusion association model, and generating an optimized spectral feature set and an optimized texture feature set; S4. Update the current image classification logic based on the optimized spectral feature set, optimized texture feature set, and crop classification model set to generate a target classification scheme; Calibrate the spatial relationships of the feature extraction steps in the target classification scheme; The S2 comprises the following steps: S21. Setting multiple feature parameter categories, including a normalized vegetation index threshold, a spatial texture complexity coefficient, and a surface temperature range, to form a feature parameter category set; Based on the crop type set and characteristic parameter category set, historical texture feature parameters, vegetation index combinations and corresponding crop recognition accuracy data are obtained to construct historical feature matrix, historical vegetation matrix and historical accuracy dataset; S22, using the historical feature matrix, the historical vegetation matrix, and the historical precision data set to train an initial feature fusion association model to obtain a feature fusion association model; The S22 includes the following steps: S221, establishing an initial feature fusion association model based on the crop type set and the feature parameter category set; S222, inputting the historical feature matrix and the historical vegetation matrix into the initial feature fusion association model row by row to generate a historical accuracy prediction data set; S223. Setting an accuracy error threshold; calculating a cross entropy error between the historical accuracy prediction dataset and the historical accuracy dataset; When the cross entropy error exceeds the accuracy error threshold, the gradient descent algorithm is used to iteratively correct the hyperparameters of the initial feature fusion association model until the error is lower than the error threshold, thus forming a feature fusion association model; The S3 includes the following steps: S31, obtaining actual spectral characteristic parameters and texture characteristic parameters of each block area in the current satellite image, and generating a real-time spectrum set and a real-time texture set; S32, setting a classification accuracy target value; inputting the real-time spectrum set and the real-time texture set into the feature fusion association model to calculate the current classification accuracy prediction value; When the current classification accuracy prediction value does not reach the target value, the real-time spectrum set and the real-time texture set are feature enhanced until the target conditions are met, and the optimized spectrum feature set and the optimized texture feature set are output.

2. The crop classification method based on multi-source satellite images according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Define multiple crop types to form a crop type set; collect corresponding near-infrared band, red-edge band, and short-wave infrared band data according to the crop type set to generate a spectral feature matrix; S12. Input the spectral feature matrix into the classification model generation module, and construct a crop classification model set through a preset feature selection algorithm.

3. The crop classification method based on multi-source satellite images according to claim 1, characterized in that: The hyperparameter correction process in S223 adopts a dynamic learning rate adjustment mechanism.

4. The crop classification method based on multi-source satellite images according to claim 1, characterized in that: Feature enhancement in S32 includes the following steps: S321. Create a feature optimization population and define the maximum evolutionary generation and the current evolutionary generation; S322, setting the adjustment range of the spectral characteristic parameters and the texture characteristic parameters to generate an initial population position matrix; S323, constructing a population fitness function based on the current classification accuracy prediction value; S324, updating the population position through multiple rounds of evolution, and recalculating the fitness value and the global optimal solution after each round of evolution; S325. When the evolutionary generation reaches the upper limit or the fitness value meets the convergence condition, the global optimal solution is extracted as the optimization result.

5. The crop classification method based on multi-source satellite images according to claim 4, characterized in that: The S4 comprises the following steps: S41, loading the optimized spectral feature set and texture feature set into the image classification unit to generate a target classification scheme; S42. Identify the spatial coordinates of the feature matching sequence in the target classification scheme, and perform geometric calibration according to a preset spatial resolution constraint.

6. The crop classification method based on multi-source satellite images according to claim 5, characterized in that: The geometric calibration in S42 uses a coordinate affine transformation algorithm, which dynamically adjusts feature matching parameters according to satellite positioning accuracy and image distortion coefficients.

7. The crop classification method based on multi-source satellite images according to claim 6, characterized in that: Also includes: S5. Obtain crop growth cycle data of images of different time phases in the target classification scheme and establish temporal feature association rules; Optimize the spatial distribution map of crop types in the classification scheme based on association rules of temporal characteristics.

Images