Crop classification method based on multi-source satellite image

Abstract

The invention relates to the technical field of satellite remote sensing application, and discloses a crop classification method based on a multi-source satellite image, and the method comprises the steps: firstly obtaining multi-temporal satellite image data, extracting spectral reflectivity characteristics, generating a matrix, and constructing a classification model set; setting a feature parameter category set, and establishing a feature fusion correlation model based on the obtained texture, vegetation index parameters and spatial resolution data of the historical image; extracting spectrum and texture feature parameters of a crop area in the current image, and generating an optimized feature set in combination with model optimization; updating classification logic based on the optimized feature set and the classification model, generating a target classification scheme and calibrating a spatial relationship; and finally, acquiring crop growth cycle data, and establishing a time sequence feature association rule to optimize a crop type spatial distribution map. According to the method, through multi-source feature fusion and dynamic optimization, the crop classification precision and reliability are improved, and the method is suitable for a precision agricultural management scene.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite remote sensing applications, specifically a method for classifying crops based on multi-source satellite images. Background Art

[0002] With the acceleration of the agricultural modernization process, the demand for rapid and accurate identification of crop types in precision agricultural management is becoming increasingly urgent. Traditional crop classification methods mainly rely on ground surveys or single-sensor data, which have problems such as limited coverage, poor timeliness, high labor costs, and being significantly affected by weather conditions, making it difficult to meet the needs of large-scale and dynamic monitoring.

[0003] Satellite remote sensing technology has become an important means for crop classification due to its advantages of macroscopic, dynamic, rapid, and multi-temporal observations. However, the existing crop classification technologies based on satellite images still face many challenges. On the one hand, the information obtained by a single satellite sensor is limited. When only using spectral features for classification, it is easily interfered by the phenomena of same object with different spectra and different objects with the same spectrum, resulting in insufficient classification accuracy. For example, the same crop in different growth stages may have different spectral reflectivities, while some different types of crops may exhibit similar spectral characteristics in specific bands, causing classification confusion.

[0004] On the other hand, the existing methods do not make full use of the fusion of multi-source satellite image data. Multi-source satellite images contain different band information (such as near-infrared, red edge, short-wave infrared, etc.), texture features, vegetation indices, and spatial resolution data, etc. How to effectively fuse these multi-dimensional features and establish an accurate classification model is the key issue for improving crop classification accuracy. In addition, when traditional classification models process multi-temporal data, they lack the effective use of the temporal characteristics of the crop growth cycle and are difficult to dynamically reflect the growth and change process of crops, resulting in insufficient continuity and reliability of the classification results in the time dimension.

[0005] At the same time, in the feature extraction and model training process, the existing technologies often adopt fixed feature selection algorithms and static model parameter settings, which cannot be adaptively adjusted according to the characteristics of image data in different regions and different time phases, resulting in poor generalization ability of the model. For example, the surface cover types in different regions are complex and diverse, and factors such as soil background and climate conditions will affect the spectral and texture characteristics of crops. Fixed model parameters are difficult to achieve ideal classification effects in various complex environments.

[0006] In addition, during the acquisition and transmission of satellite images, they may be affected by factors such as atmospheric disturbances and sensor noise, resulting in geometric distortion and radiation errors in the images, thereby affecting the accuracy of feature extraction and the reliability of classification results. Most of the existing geometric calibration methods are based on fixed transformation models and do not fully consider the dynamic changes of satellite positioning accuracy and image distortion coefficients, so the 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 purpose, the present invention provides the following technical solutions: A crop classification method based on multi-source satellite images, the method includes: S1. Obtain multi-temporal satellite image data, extract spectral reflectance characteristics of different bands, and generate a spectral feature matrix; Construct a crop classification model set according to the spectral feature matrix; S2. Set the characteristic parameter categories corresponding to various land cover types to form a characteristic parameter category set; Based on the characteristic parameter category set, obtain texture characteristic parameters, vegetation index parameters and spatial resolution data in multiple groups of historical image data, and establish a feature fusion correlation model; S3. According to the characteristic parameter category set, extract the spectral characteristic parameters and texture characteristic parameters of each crop area in the current satellite image, and combine the feature fusion correlation model to optimize the features, 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, the optimized texture feature set and the crop classification model set to generate a target classification scheme; calibrate the spatial relationship in the feature extraction link of the target classification scheme.

[0009] Preferably, the S1 includes the following steps: S11. Define multiple crop types to form a crop type set; collect data of corresponding near-infrared bands, red-edge bands and short-wave infrared bands according to the crop type set, and 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.

[0010] Preferably, the S2 includes the following steps: S21. Set multiple characteristic parameter categories, including the normalized vegetation index threshold, the spatial texture complexity coefficient and the surface temperature range, to form a characteristic parameter category set; Obtain historical texture feature parameters, vegetation index combinations, and corresponding crop recognition accuracy data based on the crop type set and the feature parameter category set, and construct a historical feature matrix, a historical vegetation matrix, and a historical accuracy data set; S22. Train the initial feature fusion association model using the historical feature matrix, the historical vegetation matrix, and the historical accuracy data set to obtain the feature fusion association model.

[0011] Preferably, S22 includes the following steps: S221. Establish an initial feature fusion association model according to the crop type set and the feature parameter category set; 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 data set; S223. Set the accuracy error threshold; calculate the cross-entropy error between the historical accuracy prediction data set and the historical accuracy data set; When the cross-entropy error exceeds the accuracy error threshold, use the gradient descent algorithm to iteratively correct the hyperparameters of the initial feature fusion association model until the error is lower than the error threshold to form the feature fusion association model.

[0012] Preferably, a dynamic learning rate adjustment mechanism is adopted in the hyperparameter correction process of S223.

[0013] Preferably, S3 includes the following steps: S31. Obtain the actual spectral feature parameters and texture feature parameters of each plot area in the current satellite image to generate a real-time spectral set and a real-time texture set; S32. Set the classification accuracy target value; input the real-time spectral 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, perform feature enhancement on the real-time spectral set and the real-time texture set until the target condition is met, and output the optimized spectral feature set and the optimized texture feature set.

[0014] Preferably, the feature enhancement in S32 includes the following steps: S321. Create a feature optimization population, and define the maximum number of evolution generations and the current number of evolution generations; S322. Set the adjustment range of the spectral feature parameters and the texture feature parameters to generate an initial population position matrix; S323. Construct a population fitness function based on the current classification accuracy prediction value; S324. Update the population position through multiple rounds of evolution, and recalculate the fitness value and the global optimal solution after each round of evolution; S325. When the number of evolution generations reaches the upper limit or the fitness value meets the convergence condition, extract the global optimal solution as the optimization result.

[0015] Preferably, S4 includes the following steps: S41. Load 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 the preset spatial resolution constraint conditions.

[0016] Preferably, in S42, the geometric calibration 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.

[0017] Preferably, the method further includes: S5. Obtain the crop growth cycle data of the images at different time phases in the target classification scheme, and establish a temporal feature association rule; optimize the spatial distribution map of the crop types in the classification scheme according to the temporal feature association rule.

[0018] Compared with the prior art, the beneficial effects of the present invention are: In terms of multi-source feature fusion, the method not only extracts the spectral reflectance features of different bands of satellite images, but also fuses multi-dimensional information such as texture feature parameters, vegetation index parameters, and spatial resolution data. By constructing a feature fusion association model, the dynamic optimization of spectral and texture features is realized. For example, in the feature enhancement process, the evolutionary algorithm is used to iteratively optimize the spectral and texture feature parameters, which can effectively solve the problems of same object with different spectra and different objects with the same spectrum, and significantly improve the classification accuracy. Compared with the traditional method that only relies on spectral features, the present invention can more comprehensively describe the feature differences of crops through the synergistic effect of multi-source features, thereby improving the accuracy of classification.

[0019] 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, and combining a preset feature selection algorithm to construct a set of crop classification models, the models can make full use of multi-band spectral information for classification. At the same time, the training process of the feature fusion association model adopts a dynamic learning rate adjustment mechanism and a gradient descent algorithm for hyperparameter optimization, which can automatically adjust the model parameters according to historical data, and improve the generalization ability and adaptive performance of the model. This dynamic optimization mechanism enables the model to better adapt to the characteristics of satellite image data in different regions and at different time phases, and still maintain a high classification accuracy in complex environments.

[0020] In the feature optimization and processing stage, through the input of real-time spectral sets and real-time texture sets, combined with the target value of classification accuracy, feature enhancement is carried out. The evolutionary algorithm is used to create a feature optimization population and conduct multiple rounds of evolution, and finally the global optimal solution is extracted as the optimization result. This feature enhancement technology based on the optimization algorithm can automatically adjust feature parameters according to the characteristics of the current image data, further improve the discriminant ability of features, and ensure that the classification results meet the accuracy requirements. Compared with the traditional feature processing method with fixed parameters, the dynamic feature optimization mechanism of the present invention can more flexibly adapt to different data conditions and improve the reliability of classification.

[0021] In terms of geometric calibration and spatial relationship processing, the coordinate affine transformation algorithm is used to calibrate the spatial relationship in the target classification scheme. This algorithm can dynamically adjust feature matching parameters according to satellite positioning accuracy and image distortion coefficients, effectively correct geometric distortion and radiation errors of satellite images, and improve the accuracy of feature extraction and the spatial positioning accuracy of classification results. By accurately 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 division and dynamic monitoring of crop planting areas.

[0022] In addition, the present invention also introduces a temporal feature association rule. By obtaining crop growth cycle data of images at different time phases, a temporal feature association rule is established and the spatial distribution map of crop types in the classification scheme is optimized. This innovation enables the classification method to make full use of the time dimension information in the crop growth process, dynamically reflect the growth changes of crops, improve the continuity and reliability of classification results in the time series, and provide more comprehensive information support for the dynamic management and decision-making of agricultural production. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is the working principle diagram of the crop classification method based on multi-source satellite images according to the present invention; Figure 2 is the design diagram of the feature fusion association model training; Figure 3 is the design diagram of the model iterative optimization; Figure 4 is the design diagram of the feature enhancement evolutionary algorithm. DETAILED DESCRIPTION OF THE INVENTION

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] Please refer to Figures 1-4 , the crop classification method based on multi-source satellite images involved in the present invention is specifically implemented as follows: S1: Obtain multi-temporal satellite image data and construct a classification model set: Obtain multi-temporal satellite image data covering the target area, extract spectral reflectance features of different bands, and form a spectral feature matrix containing spectral feature parameters. Based on this spectral feature matrix, construct a crop classification model set through a specific algorithm to provide a model basis for subsequent classification.

[0026] S2: Set the feature parameter categories and establish a feature fusion correlation model: Set the feature parameter categories corresponding to various land cover types to form a set of feature parameter categories. Based on this category set, extract texture feature parameters, vegetation index parameters, and spatial resolution data from multiple groups of historical image data, and establish a feature fusion correlation model capable of fusing multi-dimensional features through data association and model training.

[0027] S3: Extract the current image features and optimize them: According to the set of feature parameter categories, extract the spectral feature parameters and texture feature parameters of each crop area from the current satellite image, and optimize the features by combining the established feature fusion correlation model to generate an optimized spectral feature set and texture feature set, improving the classification effectiveness of the feature data.

[0028] S4: Update the classification logic and calibrate the spatial relationship: Based on the optimized spectral feature set, texture feature set, and crop classification model set, update the classification logic of the current image to generate a target classification scheme. For the spatial relationship in the feature extraction link of the target classification scheme, calibrate it through a specific algorithm to ensure the spatial position accuracy of the classification result.

[0029] The present invention will be further described below in conjunction with Embodiments 1 to 5: Embodiment 1: This embodiment elaborates on Step S1 in detail. S1 includes the processes of defining a set of crop types, collecting data of specific bands to generate a spectral feature matrix, and constructing a crop classification model set through a feature selection algorithm, which are specifically as follows: Execute Step S11: Define multiple crop types and form a set of crop types. In actual operation, specific crop types need to be determined according to 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 maize, and cotton; if it is the hilly area in the south, rice, rapeseed, tea trees, etc. can be included. The definition of the set of crop types needs to cover all the main planting types in the target area to ensure the comprehensiveness of the classification objects. After the definition is completed, a unique identification category needs to be established for each crop for subsequent data association and model training.

[0030] According to the crop type set, the corresponding near-infrared band, red-edge band and short-wave infrared band data are collected. Different bands of satellite images carry different ground information. The near-infrared band is sensitive to the chlorophyll content and leaf structure of vegetation. The red-edge band (usually in the range of 680-760nm) is the mutation area of vegetation spectral reflectance from visible light to near-infrared, which can reflect the growth status and species differences of vegetation. The short-wave infrared band (such as 1400-1900nm, 2000-2500nm) is related to the moisture content and biochemical components of vegetation. For each type of crop, it is necessary to collect satellite image data covering the target area at multiple phases in its growth cycle (such as seedling stage, flowering stage, maturity stage, etc.) to ensure that the spectral characteristics of different growth stages are obtained.

[0031] During data collection, attention should be paid to the selection of satellite sensors and data preprocessing. For example, multi-source satellite data such as the Landsat series satellites (equipped with OLI sensors, including near-infrared and short-wave infrared bands) and Sentinel-2 satellites (providing 13 bands, including red-edge bands) can be used to meet the needs of multi-band collection. The collected raw image data needs to be preprocessed by radiometric calibration, atmospheric correction, geometric correction, etc. to eliminate the influence of sensor noise, atmospheric scattering, terrain undulations and other factors on spectral reflectance, and ensure that the extracted spectral characteristic parameters are accurate and comparable.

[0032] After preprocessing, the spectral reflectance values of each pixel in the near-infrared band, red-edge band, and short-wave infrared band are extracted for each phase image of each crop type. Taking the pixel as the basic unit, the reflectance values of the three bands are combined into a feature vector, which contains the spectral feature information of the object corresponding to the pixel. For example, for a certain pixel, its near-infrared band reflectance is 0.45, the red-edge band reflectance is 0.38, and the short-wave infrared band reflectance is 0.22, then a feature vector [0.45, 0.38, 0.22] is formed. All pixel feature vectors belonging to the same crop type are summarized and combined with their corresponding crop category labels to construct a spectral feature subset of the crop type.

[0033] The spectral feature subsets of all crop types are integrated to form a global spectral feature matrix. The rows of the matrix represent different pixel samples, and the columns represent the reflectance characteristics of the three bands of near infrared, red edge, and short-wave infrared. 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 usually stored in a two-dimensional table, where the sample label column clearly identifies the crop type to which each pixel belongs.

[0034] After the spectral feature matrix is constructed, step S12 is entered: The spectral feature matrix is input into the classification model generation module, and a set of crop classification models is constructed through a preset feature selection algorithm. The classification model generation module can be an integrated learning platform or a custom algorithm framework. Its core function is to receive the spectral feature matrix, screen out the most discriminative feature combinations for classification through the feature selection algorithm, and construct multiple models based on different classification algorithms.

[0035] The role of the preset feature selection algorithm is to screen out the features with high contribution to crop classification from the spectral features of three bands, reduce the dimensionality of the feature space, avoid the "curse of dimensionality", and improve the model efficiency. Common feature selection algorithms include recursive feature elimination (RFE), random forest feature importance ranking, variance selection method, etc. Taking the recursive feature elimination algorithm as an example, its basic process is: First, train an initial classification model (such as support vector machine SVM) using all three-band features, calculate the importance score of each feature, then remove the feature with the lowest score, retrain the model on the remaining features, and repeat this process until the preset number of features is reached or the stop condition is met. Through this process, the optimal feature combination can be determined. For example, the combination of the near-infrared band and the red-edge band may be selected as the key features.

[0036] Based on feature selection, multiple crop classification models are constructed using different classification algorithms. Optional classification algorithms include support vector machine (SVM), random forest (RandomForest), gradient boosting decision tree (GBDT), neural network (such as multi-layer perceptron MLP), etc. Each algorithm has different hypothesis spaces and learning abilities. By constructing a model set, complementary advantages can be realized, and the robustness and accuracy of classification can be improved.

[0037] Taking the random forest model as an example, its construction process is: Randomly draw multiple sample subsets (bootstrap sampling) from the spectral feature matrix, randomly select some features (such as the near-infrared and red-edge bands after feature selection) on each sample subset to construct decision trees, and determine the final classification result through the voting mechanism of multiple decision trees. The random forest model has good robustness to noise data and overfitting, and is suitable for processing high-dimensional feature data.

[0038] For the support vector machine model, it is necessary to first standardize the spectral feature matrix, scale each feature value to the same range (such as [0,1] or [-1,1]), and then map the features to a high-dimensional space through a kernel function (such as radial basis kernel function RBF), and find the hyperplane that maximizes the classification margin to achieve the classification of different crop types.

[0039] When constructing a set of crop classification models, each model needs to be trained, validated, and optimized. Specifically, the spectral feature matrix is divided into a training set and a validation set (e.g., in a 7:3 ratio). The training set is used to learn the parameters of the model, and the classification accuracy, recall rate, F1 value, and other metrics of the model are evaluated through the validation set. The hyperparameters of the model (such as the number of trees in a random forest, the penalty coefficient C and kernel function parameter γ of SVM, etc.) are adjusted until the performance of the model on the validation set reaches a stable state.

[0040] Through the above process, a set of crop classification models constructed by multiple different algorithms is finally formed. Each model corresponds to different feature combinations or algorithm parameters and can model the spectral features of crops from different perspectives. The advantage of the model set is that the prediction results of multiple models can be integrated through ensemble learning methods (such as voting method, stacking method) to further improve the classification performance.

[0041] Example 2: This example elaborates on step S2 in detail. S2 includes the process of setting the feature parameter category set, obtaining multi-dimensional feature parameters in historical image data, and establishing a feature fusion association model through model training, as follows: Execute step S21: Set multiple feature parameter categories and form a feature parameter category set. The feature parameter categories need to cover the key indicators that can reflect the differences in land cover types, specifically including the normalized vegetation index threshold, spatial texture complexity coefficient, and surface temperature range. The normalized vegetation index (such as NDVI, EVI) can quantify the vegetation coverage and growth status, the spatial texture complexity coefficient is used to describe the uniformity or roughness of the surface texture (such as the contrast, entropy value, etc. calculated through the gray-level co-occurrence matrix), and the surface temperature range is related to vegetation transpiration and energy balance. The selection of these parameter categories needs to be based on the prior knowledge of crop spectral and texture features to ensure their effectiveness for crop classification.

[0042] After forming the feature parameter category set, based on the crop type set and the feature parameter category set, texture feature parameters, vegetation index parameters, and spatial resolution data are obtained from multiple groups of historical image data. The historical image data needs to cover multiple years and seasons of the target area and contain images of different crop types at different growth stages. For example, for the winter wheat and summer maize rotation area, image data at multiple time phases such as sowing period, tillering period, jointing period, and filling period need to be collected to ensure coverage of the characteristic changes of the crop throughout the growth cycle.

[0043] For the extraction of texture feature parameters, first preprocess the historical images (such as noise removal and image enhancement), and then use texture analysis algorithms (such as Gray Level Co-occurrence Matrix GLCM and Local Binary Pattern LBP) to calculate the texture features of each pixel or region. Taking the Gray Level Co-occurrence Matrix as an example, by statistically analyzing the gray level co-occurrence frequency of pixels in a specific direction and distance in the image, parameters such as contrast (reflecting the clarity of the texture), entropy (reflecting the complexity of the texture), and angular second moment (reflecting the uniformity of the texture) can be obtained. For each time-phase image of each crop type, extract its texture feature parameters and form a texture feature vector.

[0044] The extraction of vegetation index parameters needs to be based on the multi-spectral band data of historical images. Taking NDVI as an example, its calculation formula is (near-infrared band reflectance - red band reflectance) / (near-infrared band reflectance + red band reflectance), which can reflect the chlorophyll content and photosynthesis intensity of vegetation. For each crop type, calculate its NDVI, EVI and other vegetation index values in different time-phase images to form vegetation index combined data. In addition, the spatial resolution data needs to record the ground pixel resolution of the corresponding image (such as 10 meters, 20 meters) for scale matching in subsequent feature fusion.

[0045] While obtaining the above data, it is necessary to collect the corresponding crop recognition accuracy data. This data can be obtained through on-site actual surveys, agricultural statistical data or manual interpretation, etc., and is used to label the actual crop types in each region of the historical images as the true value labels for model training. For example, obtain the sample point coordinates through on-site sampling, combine with the image data to determine the crop category corresponding to the sample point, and establish a mapping relationship of "feature parameter - crop type".

[0046] Based on the above data, construct a historical feature matrix, a historical vegetation matrix and a historical accuracy data set. The rows of the historical feature matrix represent samples (such as plots or pixels in the image), and the columns correspond to texture feature parameters (such as contrast, entropy value, etc.). Each row records the texture feature vector of a sample and the corresponding crop category label; the historical vegetation matrix stores the vegetation index combined data (such as NDVI, EVI values) of each sample and the category label; the historical accuracy data set contains the classification accuracy reference values of each sample under a specific feature combination, and this value can be calculated by comparing the model prediction results with the true value labels (such as the comprehensive evaluation value of indicators such as accuracy rate and recall rate).

[0047] After completing the data construction, enter step S22: Use the historical feature matrix, the historical vegetation matrix and the historical accuracy data set to train the initial feature fusion association model to obtain the feature fusion association model. This process realizes the fusion of multi-dimensional features and the prediction of classification accuracy through machine learning algorithms. The specific steps are as follows: Execute step S221: Establish an initial feature fusion association model based on the crop type set and the feature parameter category set. The initial model can adopt architectures such as support vector machine (SVM), neural network (such as multi-layer perceptron MLP), or ensemble learning model (such as gradient boosting machine GBM). Taking the 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 feature contains 4 parameters and the vegetation index contains 2 parameters, then the input layer has 6 nodes). The hidden layer realizes the non-linear transformation of features through an activation function (such as ReLU), and the number of nodes in the output layer corresponds to the number of categories of crop types, and the classification probability is output through the Softmax function.

[0048] 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 data set. Before inputting the data, it is necessary to standardize the texture features and vegetation indices (such as normalizing them to the [0,1] interval) to avoid the impact of different feature scale differences on model training. The model calculates through forward propagation, fuses the texture and vegetation index features of each sample, and outputs the classification accuracy prediction value of the sample, which reflects the model's estimation of the classification accuracy under the feature combination of the sample.

[0049] Execute step S223: Set the accuracy error threshold and calculate the cross-entropy error between the historical accuracy prediction data set and the historical accuracy data set. The accuracy error threshold is a standard for measuring whether the model prediction accuracy meets the standard, and it needs to be set according to the actual needs of the classification task (such as 0.1 or 0.2). The cross-entropy error is used to quantify the difference between the predicted value and the true value, and its calculation formula is to take the average after summing the logarithmic losses of the true accuracy values and predicted accuracy values of all samples.

[0050] When the cross-entropy error exceeds the accuracy error threshold, trigger the model parameter correction mechanism, and use the gradient descent algorithm 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 along the opposite direction of the gradient to minimize the error. During the iteration process, the following details need to be noted: Batch processing: Stochastic gradient descent (SGD) or mini-batch gradient descent (Mini-BatchGD) can be adopted to divide the training data into multiple batches and update the parameters step by step to avoid excessive memory occupation and low computational efficiency.

[0051] Learning rate control: The learning rate determines the step size of parameter update. The initial learning rate can be set to a small value such as 0.01 to ensure the stability of parameter update. As the iteration progresses, the learning rate can be gradually reduced through a learning rate decay strategy (such as multiplying the learning rate by 0.9 every 50 iterations) to avoid parameter oscillation in the later stage.

[0052] Regularization: To prevent the model from overfitting, L1 or L2 regularization terms can be added to the loss function to penalize the complexity of the model parameters and improve the generalization ability of the model.

[0053] During the iterative process, it is necessary to regularly evaluate the change trend of the cross-entropy error. When the decrease in the error for consecutive multiple rounds of iteration is less than the preset threshold (such as 0.001) or reaches the maximum number of iterations (such as 1000 rounds), it is considered that the model converges, and the iteration stops, finally forming a feature fusion correlation model. This model can effectively fuse texture features, vegetation index features, and spatial resolution information, establish a mapping relationship between multi-dimensional features and crop classification accuracy, and provide support for subsequent feature optimization of the current image.

[0054] During the model training process, attention should also be paid to the rationality of data division. Usually, historical data is divided into a training set (about 70%), a validation set (about 20%), and a test set (about 10%). The training set is used for model parameter learning, the validation set is used to adjust hyperparameters and monitor overfitting, and the test set is used to evaluate the generalization ability of the model. Through cross-validation (such as k-fold cross-validation), the reliability of model evaluation can be further improved to ensure that the feature fusion correlation model can show stable performance on different data subsets.

[0055] Example 3: This example elaborates in detail on the hyperparameter correction process in step S223, focusing on the application of the dynamic learning rate adjustment mechanism in the gradient descent algorithm.

[0056] During the training process of the feature fusion correlation model, when the cross-entropy error exceeds the accuracy error threshold, it is necessary to iteratively correct the hyperparameters of the model (such as the weights, biases, or regularization coefficients of the neural network) through the gradient descent algorithm. The traditional gradient descent algorithm with a fixed learning rate may face problems such as slow initial convergence speed or parameter oscillation in the later stage. Therefore, a dynamic learning rate adjustment mechanism is introduced, and its core idea is to automatically adjust the learning rate according to the training process to balance the convergence speed and optimization accuracy.

[0057] The specific implementation method 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 , and the maximum number of iterations is , then the learning rate can be expressed as:

[0058] Among them, is the learning rate decay function, which is used to describe the change law of the learning rate with the number of iterations. Common decay functions include exponential decay, polynomial decay, and cosine decay, etc. The implementation details are elaborated below respectively.

[0059] Exponential decay: The expression of the exponential decay function is:

[0060] where is the decay rate ( ), which needs to be preset according to the model training situation. For example, if , the learning rate decays to 95% of the previous value after each iteration. The advantage of this method is its simplicity in implementation, which can quickly reduce the learning rate to avoid late-stage oscillations. However, the decay speed is fixed and may not be able to adapt to complex training curves.

[0061] In practical applications, exponential decay can be combined with a warm-up mechanism. That is, in the initial stage of training (such as the first iterations), the learning rate does not decay immediately but increases linearly to to help the model quickly explore the parameter space in the initial stage. The expression of the learning rate in the warm-up stage is:

[0062] After the warm-up ends, it enters the exponential decay stage. This method is suitable for scenarios where the model parameter initialization is poor and can improve the initial convergence stability.

[0063] Polynomial decay: The polynomial decay function describes the decreasing trend of the learning rate through a polynomial function, and its general form is:

[0064] where is the degree of the polynomial (a positive integer). When , it degenerates to linear decay; when , the learning rate drops rapidly in the initial stage and flattens out in the later stage. For example, if , the learning rate decays to 25% of the initial value in the middle of the iteration ( ), which is suitable for models that require fast initial convergence and fine-tuning in the later stage.

[0065] The advantage of polynomial decay is that the decay speed can be flexibly controlled by adjusting the value, but the maximum number of iterations needs to be determined in advance. If the training converges early in practice, it may cause waste of computing resources. Therefore, an early stopping mechanism (EarlyStopping) can be combined. When the validation set error no longer decreases, the training is terminated early to avoid ineffective iterations.

[0066] Cosine decay: The cosine decay function simulates the decreasing trend of the cosine curve, and its expression is:

[0067] This function decreases slowly at the beginning, accelerates in the middle, slows down again later, and finally approaches 0. The characteristic of cosine decay is that the learning rate will oscillate slightly near the minimum value in the later stage of training, which may help the model jump out of local optima and improve generalization ability. For example, in the later stage of training, small fluctuations in the learning rate can enable more refined adjustment of parameters near the optimal solution.

[0068] To further enhance the flexibility of cosine decay, periodic cosine decay (CosineAnnealing) can be introduced, that is, after the learning rate drops to a certain level, it resumes to a higher value and decays again, forming multiple decay cycles. The maximum number of iterations in each cycle can be decreased in a certain proportion (such as halving the cycle length each time) to simulate the annealing process and help the model search for the optimal solution at different scales.

[0069] In the implementation process of the dynamic learning rate adjustment mechanism, the following key aspects need to be noted: Learning rate monitoring: Record the current learning rate and the corresponding cross-entropy error after each iteration, plot the learning rate-error curve, and observe the relationship between the two. If it is found that the error continues to rise after the learning rate drops, it may indicate that the learning rate decays too fast, and the parameters of the decay function need to be adjusted (such as increasing or decreasing ).

[0070] Parameter synchronous update: In deep learning frameworks (such as TensorFlow, PyTorch), dynamic learning rate adjustment is usually achieved through optimizers. For example, the torch.optim.lr_scheduler module in PyTorch provides various learning rate schedulers (such as ExponentialLR, CosineAnnealingLR), which can automatically update the learning rate according to the number of iterations and ensure the synchronous adjustment of model parameters and the learning rate.

[0071] Hyperparameter co-optimization: The parameters in the dynamic learning rate adjustment mechanism (such as , , ) need to be co-optimized with other hyperparameters of the model (such as the number of neurons in the hidden layer, regularization coefficient). Methods such as grid search (GridSearch) or random search (RandomSearch) can be used to find the optimal parameter combination on the validation set.

[0072] Taking a neural network model as an example, assume the initial learning rate , using the exponential decay function, the decay rate , and the maximum number of iterations . In the initial stage of training (the first 100 iterations), the learning rate gradually decays from 0.01 to approximately ; In the middle stage of training (the 500th iteration), the learning rate drops to approximately ; In the later stage 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 finally converges to a better parameter space.

[0073] The introduction of the dynamic learning rate adjustment mechanism makes the hyperparameter correction process more adaptable and intelligent. Compared with a fixed learning rate, its advantages are reflected in: Fast convergence in the initial stage: A higher learning rate can accelerate the update of parameters in the initial stage and shorten the training time.

[0074] Stable optimization in the later stage: The decayed learning rate can prevent the parameters from oscillating back and forth near the optimal solution and improve the convergence accuracy.

[0075] Enhanced robustness: Through adaptive adjustment, it can cope with the differences in feature distributions of different datasets and reduce the cost of manual hyperparameter tuning.

[0076] In specific implementation, it is necessary to select a suitable decay strategy according to factors such as the data scale of the historical feature matrix and historical vegetation matrix, and the model complexity. For example, for a small dataset or a simple model, linear decay or exponential decay may be sufficient; for a large dataset or a deep neural network, cosine decay or periodic decay may be more effective.

[0077] Example 4: This example elaborates on step S3 in detail, and illustrates the whole process of extracting features, optimizing features, and generating an optimized feature set from the current satellite image in a specific application scenario as follows: Suppose the target area is an agricultural demonstration area in a certain province, and currently it is necessary to classify crops in the Sentinel-2 satellite image of this area in May 2025. First, execute step S31: Obtain the actual spectral feature parameters and texture feature parameters of each plot area in the current satellite image, and generate a real-time spectral set and a real-time texture set. In actual operation, load the image data through remote sensing image processing software (such as ENVI, QGIS), and use a vector boundary file (such as the plot distribution map provided by the agricultural department) to divide the image into different plot units. For example, for plot No. A101, extract the reflectance values in the blue, green, red, near-infrared and other bands as spectral feature parameters; at the same time, use the gray-level co-occurrence matrix algorithm to calculate the texture feature parameters of this plot, including contrast, entropy value, homogeneity, etc., which reflect the planting uniformity or leaf surface structure differences of the crops in the plot. Repeat the above operations for all plots to form a real-time spectral set (recording the reflectance of each band) and a real-time texture set (recording the texture feature parameters) containing thousands of plot samples, and each sample is associated with its plot number and geographical location coordinates.

[0078] Next, enter step S32: Set the classification accuracy target value and start the feature optimization process. Assume that the classification accuracy target value is set to 90% according to business requirements. Input the real-time spectral set and real-time texture set into the trained feature fusion and association model (such as the neural network model established in Embodiment 2). The model calculates the current classification accuracy prediction value through forward propagation. If the prediction value is 85% (not reaching the target value), then trigger the feature enhancement process, which is specifically implemented through the following steps: S321: Create a feature optimization population and initialize parameters. Define the feature optimization population as a set of potential feature adjustment schemes, where each individual represents a combination of adjustments to spectral feature parameters and texture feature parameters. For example, the population size is set to 50 individuals, the maximum number of evolution generations is 100 generations, and the current evolution generation is initialized to 0. The gene encoding of each individual corresponds to the adjustment amplitude of spectral feature parameters (such as near-infrared band reflectance) and texture feature parameters (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.).

[0079] S322: Set the feature adjustment range and generate the initial population position matrix. According to the physical meaning of spectral and texture features and the historical data distribution, set the adjustment range of each feature parameter. For example, the adjustment range of near-infrared band reflectance is ±10% (to avoid exceeding the normal reflectance range of vegetation), and the adjustment range of texture contrast is ±15% (considering the texture change amplitude in the crop growth stage). Generate the initial population position matrix through uniform random sampling, where each row vector corresponds to the feature adjustment scheme of an individual. For example, the adjustment scheme of the first individual may be: near-infrared reflectance +8%, red-edge band reflectance -5%, texture entropy value +10%, and the second individual is near-infrared reflectance -3%, red-edge band reflectance +6%, texture contrast -8%, etc., to ensure that the initial population covers different regions of the feature space.

[0080] S323: Construct the population fitness function. The fitness function aims to improve the classification accuracy and uses the prediction accuracy of the feature fusion and association model for the adjusted features as the fitness value. Specifically, for the feature adjustment scheme of each individual, adjust the real-time spectral set and real-time texture set according to this scheme (such as multiplying the near-infrared band reflectance value by 1.08 and adding 10% of the original value to the texture entropy value) to generate the adjusted feature set, and then input it into the model to calculate the new classification accuracy prediction value, which is used as the fitness value of this individual. For example, if the prediction accuracy after adjustment of an individual is 87%, its fitness value is 0.87; another individual is 89% after adjustment, and the fitness value is 0.89. The latter has higher survival competitiveness in the population.

[0081] S324: Update the population position through multiple rounds of evolution. Use the basic operations of genetic algorithms (selection, crossover, mutation) for population evolution: Selection operation: Sort the individuals in the population based on the fitness value, and use the roulette wheel selection or tournament selection method to retain the individuals with higher fitness (such as the top 30 individuals) to enter the next generation, and eliminate the individuals with lower fitness. For example, in tournament selection, 5 individuals are randomly selected each time, and the individual with the highest fitness is retained, and the process is repeated until the population size of the next generation is full.

[0082] Crossover operation: Perform pairwise crossover on the selected individuals, and exchange some gene segments to generate new individuals. For example, for the feature adjustment schemes of two individuals, a single-point crossover is performed at the 3rd gene position (such as the adjustment of the texture entropy value) to generate two offspring individuals, whose genes inherit the first half and the second half of the adjustment parameters of the parent generation respectively. The crossover probability is usually set to 0.8 - 0.9 to ensure the diversity of the population.

[0083] Mutation operation: Mutate the genes of the crossed individuals with a low probability (such as 0.01 - 0.05), and randomly change the adjustment amplitude of a certain feature to avoid the population converging to the local optimum prematurely. For example, the adjustment amplitude of the red-edge band reflectance of an individual is +6%, and after mutation, it may become +9% or -2%, so as to explore new regions of the feature space.

[0084] After each round of evolution, recalculate the fitness values of all individuals and record the global optimal solution (that is, the feature adjustment scheme with the highest current fitness). For example, after the 10th generation of evolution, the prediction accuracy corresponding to the global optimal solution is increased to 88%; at the 30th generation, it reaches 89.5%, approaching the target value.

[0085] S325: Terminate the evolution and extract the optimization results. When the number of evolution generations reaches the maximum number of generations (such as the 100th generation) or the fitness value changes less than 0.1% for multiple consecutive generations (such as 5 generations), terminate the evolution process. Extract the feature adjustment scheme corresponding to the global optimal solution and apply it to the real-time spectral set and the real-time texture set to generate the optimized spectral feature set and the optimized texture feature set. For example, the global optimal solution may be: near-infrared reflectance +7%, red-edge band reflectance -4%, texture contrast -5%, texture entropy value +12%. After adjusting through this scheme, the model prediction accuracy reaches 90.2%, meeting the target conditions.

[0086] During the feature enhancement process, the following key details need to be noted: Feature physical constraints: When adjusting the spectral feature parameters, it is necessary to ensure that the adjusted values are within the reasonable measurement range of the satellite sensor. For example, the reflectance value should be between 0 and 1 to avoid negative numbers or values exceeding 1; the adjustment of the texture feature parameters should conform to the actual texture characteristics of the ground objects (such as the texture complexity of crop leaves will not change drastically with the adjustment).

[0087] Computational efficiency optimization: When the volume of image data is large (such as containing hundreds of thousands of plots), a distributed computing framework (such as Spark) can be used to parallelize the fitness calculation of population individuals, reducing the training time. In addition, an early stopping mechanism can be set. If the fitness value has reached the target value in the middle of the evolution, the evolution can be terminated early to save computing resources.

[0088] Multi-feature collaborative adjustment: Spectral features and texture features have an inherent correlation (for example, areas with dense vegetation usually have a higher near-infrared reflectance and lower texture complexity). Therefore, when adjusting features, it is necessary to avoid excessive adjustment of a single feature, which may lead to distortion of physical meaning. For example, if the near-infrared reflectance increases significantly, it is necessary to synchronously consider whether the texture complexity conforms to the growth law of vegetation, and avoid generating unreasonable feature combinations such as "high reflectance + high texture complexity".

[0089] Taking corn and soybean plots as examples to illustrate the practical application of feature optimization: Suppose the near-infrared reflectance in the original spectral features of a plot is 0.4, the reflectance in the red-edge band is 0.35, and the texture contrast is 150 (the normalized value calculated based on the gray-level co-occurrence matrix). The initial prediction of the model is "soybean", but the actual ground survey is "corn". Through the evolution of the feature optimization population, it may be found that when the near-infrared reflectance is adjusted to 0.45 (+12.5%) and the texture contrast is adjusted to 130 (-13.3%), the model prediction result is corrected to "corn", indicating that this adjustment scheme effectively enhances the feature distinguishability between corn and soybeans.

[0090] In another scenario, the original texture entropy value of a wheat plot is 80 (reflecting the leaf structure complexity), and the NDVI value in the spectral features is 0.65. The initial prediction accuracy of the model is low. After feature optimization, the NDVI value is adjusted to 0.68 (+4.6%), and the texture entropy value is adjusted to 75 (-6.25%). The corresponding prediction accuracy is increased from 82% to 91%, indicating that by collaboratively adjusting spectral and texture features, the feature differences of wheat at the jointing stage are effectively highlighted.

[0091] Example 5: This example elaborates in detail on step S4 and the extended step S5, and illustrates the whole process of generating the target classification scheme, calibrating the spatial relationship, and optimizing the temporal features in combination with specific application scenarios, as follows: Assume that the target area is an agricultural area in a certain basin, which includes three main crops: wheat, corn, and rice. Currently, it is necessary to process Sentinel-2 satellite images (spatial resolution of 10 meters) in May 2025 and Landsat-8 satellite images (spatial resolution of 30 meters) in June 2025. First, perform step S41: Load the optimized spectral feature set and texture feature set into the image classification unit to generate the target classification scheme. The image classification unit integrates the crop classification model set (such as random forest and support vector machine models) constructed in Example 1, and uses the optimized features as input to classify and predict each plot.

[0092] Taking the plot numbered B203 as an example, its optimized spectral features include a near-infrared band reflectance of 0.48 and a red-edge band reflectance of 0.39, and its optimized texture features include a contrast of 120 and an entropy value of 75. The random forest model votes through multiple decision trees and predicts that this plot is "wheat"; the support vector machine model calculates based on the radial basis kernel function and predicts it as "wheat". Combining the output results of the model set, the target classification scheme marks this plot as "wheat". Repeat this process for all plots in the entire region to generate a preliminary classification result containing the crop types of each plot, such as "Plot B101: corn, Plot B102: rice, Plot B103: wheat", etc.

[0093] Next, perform step S42: Identify the spatial coordinates of the feature matching sequence in the target classification scheme and perform geometric calibration according to the preset spatial resolution constraint conditions. The spatial coordinates of the feature matching sequence come from the geolocation information of the satellite image. For example, the upper left corner coordinates of plot B203 are (116.5° east longitude, 39.2° north latitude), and the lower right corner coordinates are (116.501° east longitude, 39.199° north latitude). The preset spatial resolution constraint conditions are set according to 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).

[0094] The coordinate affine transformation algorithm is used for geometric calibration, and this algorithm is implemented through the following steps: Ground control point selection: Select evenly distributed ground control points (GCPs) in the study area, such as road intersections and inflection points of farmland boundaries, and obtain their true coordinates through GPS measurement ( ), and mark the corresponding pixel coordinates on the image ( ). For example, select 10 control points to cover the four corners and the center position of the area.

[0095] Transformation parameter calculation: Based on the image coordinates and true coordinates of the control points, establish an affine transformation equation system to solve for translation ( ), rotation ( ), scaling ( parameters such as (etc.) to minimize the deviation between the image coordinates and the true coordinates.

[0096] Coordinate calibration: Substitute the spatial coordinates of all plots in the target classification scheme into the affine transformation formula to generate the calibrated coordinates. For example, the original upper-left corner coordinates of plot B203 are adjusted to (116.5005° E, 39.2002° N) after calibration, which is closer to the actual field position.

[0097] The coordinate affine transformation algorithm dynamically adjusts the feature matching parameters according to the satellite positioning accuracy (such as the positioning accuracy of Sentinel-2 is 5 meters) and the image distortion coefficient (obtained through sensor parameters). For example, if the image distortion coefficient is large in a certain area due to terrain undulation, increase the number of control points and reduce the calibration window to improve the calibration accuracy of the local area; if the satellite positioning accuracy is high, appropriately reduce the control point density to improve the processing efficiency.

[0098] After completing the geometric calibration, enter the extended step S5: Obtain the crop growth cycle data of images in different time phases in the target classification scheme, establish the temporal feature association rules, and optimize the spatial distribution map of crop types. Taking winter wheat as an example, its growth cycle includes the sowing period (October), tillering period (November), green-recovery period (March), jointing period (April), heading period (May), and maturity period (June). The spectral features (such as NDVI values) and texture features (such as leaf surface complexity) of images in different time phases show regular changes with the growth stage: In the sowing period, the ground is covered with soil, the NDVI value is relatively low (about 0.2), and the texture complexity is relatively high (reflecting the rough surface of the plowed soil); in the tillering period, the vegetation coverage increases, the NDVI rises to 0.4 - 0.5, and the texture complexity decreases (the leaves start to grow evenly); in the jointing period, the NDVI can reach 0.7 - 0.8, and the texture contrast increases (the structural difference between the stem and the leaves becomes larger).

[0099] Based on the above rules, establish the temporal feature association rules: If a certain plot is classified as "wheat" in the May image, then its NDVI value in the June image should show a downward trend (entering the maturity period, the leaves start to wither), and the texture contrast should further increase (the spike structure increases the texture complexity). On the contrary, if a certain plot is classified as "corn" in May (the sowing period is usually in June), then it should show a low NDVI value (the ground is bare or covered with weeds) and a high texture complexity (the rough ground without sowing or overgrown with weeds) in the May image.

[0100] The specific process of applying the temporal feature association rules to optimize the spatial distribution map is as follows: Multi-temporal data matching: Register the current classification results (May 2025) spatially 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 are consistent in images of different temporal phases.

[0101] Temporal feature analysis: For each plot, extract spectral features (such as reflectance of each band, NDVI) and texture features (such as contrast, entropy value) of different temporal phases to generate temporal feature curves. For example, the NDVI of plot B301 was 0.3 in November 2024, 0.85 in May 2025, and 0.78 in June 2025, which conforms to the NDVI change law of winter wheat from the tillering stage to the maturity stage, and its classification result of "wheat" can be determined to be reasonable.

[0102] Abnormal classification correction: If a plot is classified as "rice" in May 2025, but its NDVI value in November 2024 is 0.7 (significantly higher than the surface characteristics of the rice fallow period), the abnormal detection mechanism is triggered. Combining historical planting pattern data (rice in this area is usually sown in spring), it is judged that this classification result may be incorrect, and the classification model is called again for review and corrected to "wheat" or other reasonable types.

[0103] Spatial distribution optimization: Adjust the classification results of adjacent plots according to the consistency constraints of temporal features. For example, in the classification results of a certain area in May 2025, there are alternating distributions of "wheat" and "corn" plots. However, according to historical temporal data, this area is a continuous wheat cropping area without corn planting records, so the corn plots are corrected to wheat to ensure that the spatial distribution map is consistent with the actual planting pattern.

[0104] Taking the classification optimization of rice and corn as an example: Suppose plot C101 is classified as "corn" in the May image, but its NDVI value in the June 2025 image is 0.9 (significantly higher than the normal range of the corn seedling stage), and the texture contrast is low (the leaves grow evenly, conforming to the characteristics of the rice tillering stage). Through the analysis of temporal feature association rules, it is found that this plot may be in the rice transplanting period in May (low NDVI) and enter the tillering stage in June (rising NDVI and uniform texture), so the classification result is corrected to "rice" to avoid misjudgment caused by single-temporal feature misjudgment.

[0105] In another scenario, plot C202 was classified as "fallow land" in May 2025, but its NDVI value in the November 2024 image was 0.6 (close to the wheat tillering stage level), and its NDVI value in the June 2025 image was 0.85 (close to the wheat maturity stage). Through the analysis of temporal feature continuity, it is judged that this plot is actually planted with wheat, and the classification result in May was misjudged due to sensor noise or insufficient feature optimization. After correction, it is included in the wheat classification area to improve the accuracy of the spatial distribution map.

[0106] It should be noted that, in this text, relational terms such as first and second are only used 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 term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0107] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present 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, It includes the following steps: S1. Obtain multi-temporal satellite image data, extract spectral reflectance features of different bands, and generate a spectral feature matrix; Construct a set of crop classification models based on the spectral feature matrix; S2. Set characteristic parameter categories corresponding to multiple land cover types to form a set of characteristic parameter categories; Based on the set of characteristic parameter categories, obtain texture feature parameters, vegetation index parameters, and spatial resolution data in multiple groups of historical image data, and establish a feature fusion correlation model; S3. Extract spectral feature parameters and texture feature parameters of each crop area in the current satellite image according to the set of characteristic parameter categories, combine with the feature fusion correlation model for feature optimization, and generate 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, the optimized texture feature set, and the set of crop classification models to generate a target classification scheme; Calibrate the spatial relationship in the feature extraction link of the target classification scheme.

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

3. The crop classification method based on multi-source satellite images according to claim 2, characterized in that, The S2 includes the following steps: S21. Set multiple characteristic parameter categories, including the normalized vegetation index threshold, the spatial texture complexity coefficient, and the surface temperature range, to form a set of characteristic parameter categories; Based on the set of crop types and the set of characteristic parameter categories, obtain historical texture feature parameters, vegetation index combinations, and corresponding crop recognition accuracy data, and construct a historical feature matrix, a historical vegetation matrix, and a historical accuracy data set; S22. Use the historical feature matrix, the historical vegetation matrix, and the historical accuracy data set to train an initial feature fusion correlation model to obtain a feature fusion correlation model.

4. The crop classification method based on multi-source satellite images according to claim 3, wherein The S22 includes the following steps: S221. Establish an initial feature fusion correlation model according to the set of crop types and the set of characteristic parameter categories; S222. Input the historical feature matrix and the historical vegetation matrix row by row into the initial feature fusion correlation model to generate a historical accuracy prediction data set; S223. Set an accuracy error threshold; calculate the cross-entropy error between the historical accuracy prediction data set and the historical accuracy data set; When the cross-entropy error exceeds the accuracy error threshold, use the gradient descent algorithm to iteratively correct the hyperparameters of the initial feature fusion correlation model until the error is lower than the error threshold to form a feature fusion correlation model.

5. The crop classification method based on multi-source satellite images according to claim 4, wherein: In the hyperparameter correction process of S223, a dynamic learning rate adjustment mechanism is adopted.

6. The crop classification method based on multi-source satellite images according to claim 5, characterized in that, The S3 includes the following steps: S31. Obtain the actual spectral feature parameters and texture feature parameters of each plot area in the current satellite image to generate a real-time spectral set and a real-time texture set; S32. Set a classification accuracy target value; input the real-time spectral set and the real-time texture set into the feature fusion correlation model to calculate the current classification accuracy prediction value; When the current classification accuracy prediction value does not reach the target value, perform feature enhancement on the real-time spectral set and the real-time texture set until the target conditions are met, and output the optimized spectral feature set and the optimized texture feature set.

7. The crop classification method based on multi-source satellite images according to claim 6, wherein The feature enhancement in S32 includes the following steps: S321. Create a feature optimization population, and define the maximum number of evolutionary generations and the current number of evolutionary generations; S322. Set the adjustment ranges of the spectral feature parameters and the texture feature parameters, and generate an initial population position matrix; S323. Construct a population fitness function based on the current classification accuracy prediction value; S324. Update the population position through multiple rounds of evolution, and recalculate the fitness value and the global optimal solution after each round of evolution; S325. When the number of evolutionary generations reaches the upper limit or the fitness value meets the convergence condition, extract the global optimal solution as the optimization result.

8. The crop classification method based on multi-source satellite images according to claim 7, characterized in that The said S4 includes the following steps: S41. Load the optimized spectral feature set and the 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 the preset spatial resolution constraint conditions.

9. The crop classification method based on multi-source satellite images according to claim 8, wherein: In S42, the coordinate affine transformation algorithm is used for geometric calibration, and the coordinate affine transformation algorithm dynamically adjusts the feature matching parameters according to the satellite positioning accuracy and the image distortion coefficient.

10. The crop classification method based on multi-source satellite images according to claim 9, characterized in that, It also includes: S5. Obtain the crop growth cycle data of the images in different time phases in the target classification scheme, and establish a temporal feature association rule; Optimize the spatial distribution map of the crop types in the classification scheme according to the temporal feature association rule.

Images