Ecological environment ground feature category rapid identification method based on convolutional neural network

Abstract

The invention discloses a convolutional neural network-based ecological environment ground object type rapid identification method. The method comprises the following identification steps of S1, performing data preprocessing on a hyperspectral remote sensing image; s2, feature extraction: extracting multi-scale spectral and spatial features through convolution and pooling operations, and comprehensively capturing local and global information of different ground feature categories; s3, training the model until the model reaches an optimal state; s4, classifying each pixel based on the spectral and spatial features, and generating a classification result graph of ground feature categories; and S5, performing post-processing on a result. By optimizing data preprocessing, automatic feature extraction and efficient result post-processing technologies, the large-scale remote sensing data processing efficiency and classification precision are improved, the calculation complexity is reduced, the generalization ability of the model in different ecological environments is enhanced, rapid and accurate recognition of complex ecological environment monitoring tasks can be achieved, and the method is suitable for popularization and application. And real-time and large-scale data processing requirements are met.

Description

Technical Field

[0001] The present invention relates to a method for quickly identifying land cover types, and particularly to a method for quickly identifying land cover types in the ecological environment based on a convolutional neural network. Background Art

[0002] In ecological environment monitoring and management, accurately identifying land cover types (such as vegetation, water bodies, bare soil, etc.) is an important part of environmental change assessment. Existing remote sensing image land cover classification methods can be mainly divided into the following three categories: rule-based classification methods, statistical machine learning methods, and convolutional neural network methods based on deep learning.

[0003] (1) Rule-based classification methods

[0004] Rule-based classification methods rely on classification rules set by expert experience, analyze the spectral features in remote sensing images, and distinguish different land cover types by manually setting thresholds. This method is suitable for simple scene classification, but when facing complex and variable remote sensing data, the rules are difficult to generalize, and it depends on manual adjustment and cannot be extended to large-scale data sets, resulting in low efficiency.

[0005] (2) Statistical machine learning methods

[0006] With the progress of machine learning technology, statistical algorithms such as support vector machines (SVM) and random forests (Random Forest) have become common remote sensing image classification tools. They train classification models through labeled data, reducing the need for manual rule design. However, these methods rely on manual feature extraction and are difficult to fully utilize multi-dimensional information (such as spectral, spatial, texture, etc.) in the images, showing limitations in processing hyperspectral and complex data.

[0007] (3) Deep learning methods based on convolutional neural networks

[0008] In recent years, CNNs have shown good performance in remote sensing image classification. Through multi-layer convolutional structures, CNNs can automatically extract multi-scale features such as spectral and spatial in the images, greatly improving the classification accuracy. Although the feature capabilities of CNNs exceed traditional methods, they still face problems such as high computational complexity, long training time, and slow recognition speed when processing large-scale remote sensing data, making it difficult to meet the requirements of real-time applications. In addition, the generalization ability of existing CNN models is limited, and when processing different data sources or new scenarios, the classification effect fluctuates greatly, affecting the stability of applications. Summary of the Invention

[0009] In order to solve the deficiencies of the above technologies, the present invention provides a method for quickly identifying land cover types in the ecological environment based on a convolutional neural network.

[0010] To solve the above technical problems, the technical solution adopted by the present invention is: a method for quickly identifying ecological environment ground object categories based on a convolutional neural network, including the following identification steps:

[0011] S1. Data preprocessing: Perform data preprocessing on the hyperspectral remote sensing image;

[0012] S2. Feature extraction: Based on the convolutional neural network, perform feature extraction on the preprocessed hyperspectral remote sensing data, and extract multi-scale spectral and spatial features through convolution and pooling operations to comprehensively capture local and global information of different ground object categories;

[0013] S3. Model training: Use the labeled training set for model learning. By setting the learning rate and hyperparameters, use the cross-entropy loss function to calculate the difference between the predicted label and the true label, and adjust the network weights through backpropagation; After multiple iterations and cross-validations, gradually reduce the loss value until the model reaches the optimal state;

[0014] S4. Category identification: After training is completed, the test set is input into the trained model for classification prediction; The model classifies each pixel based on spectral and spatial features to generate a classification result map of ground object categories;

[0015] S5. Result postprocessing: Perform postprocessing on the classification result map generated in S4. The postprocessing includes smoothing, denoising, and spatial consistency optimization, and finally output a high-quality ground object classification map.

[0016] Preferably, in S1, the preprocessing method includes dimensionality reduction, noise removal, normalization, and data augmentation.

[0017] Preferably, the principal component analysis method is used to perform dimensionality reduction on the data, calculate the covariance matrix between bands, extract the main components, compress the data dimension, and retain the key spectral information;

[0018] Use Gaussian filtering or median filtering technology to remove noise;

[0019] Scale all band data to a unified range through normalization processing;

[0020] The data augmentation process includes random rotation, scaling, and flipping operations.

[0021] Preferably, in S2, the feature extraction process is as follows:

[0022] 2.1 Establish a ResNet network architecture, which is composed of multiple residual blocks, and each residual block is composed of multiple convolutional layers and skip connections;

[0023] 2.2 Convolution operation: In the convolutional layer, use a 3×3 convolutional kernel to perform convolution operation on the input data;

[0024] 2.3 Pooling operation: After the convolution operation, the max-pooling operation is adopted to reduce the spatial dimension of the feature map while retaining the most representative features;

[0025] 2.4 Spectral and spatial feature extraction: Through the ResNet network architecture, the model captures the differences between spectral bands to distinguish different land cover classes; through multi-layer convolution operations, spatial features are extracted layer by layer to identify the shape, structure, and spatial distribution of land covers;

[0026] 2.5 Introduce residual blocks into the ResNet network structure.

[0027] Preferably, in S3, the model training process specifically includes:

[0028] 3.1 Dataset division and preprocessing: Divide into a training set and a test set, and perform data augmentation on the training set data, including random rotation, scaling, and flipping;

[0029] 3.2 Network model initialization: The initialization of the convolutional neural network is based on the ResNet network architecture and introduces residual connections; the initial weight parameters are set randomly;

[0030] 3.3 Loss function calculation: Based on the cross-entropy loss function to evaluate the difference between the predicted label and the true label;

[0031] 3.4 Error backpropagation and weight update: The loss value is propagated layer by layer through backpropagation to each layer of the network, and the optimizer updates the weight parameters according to the loss value; the optimizer controls the step size of each weight update by adjusting the learning rate;

[0032] 3.5 Multiple rounds of iterative training and dynamic parameter adjustment;

[0033] 3.6 Model validation and performance optimization: After each round of training, the generalization ability and classification accuracy of the model are verified through the test set; the model performance is gradually improved by adjusting the optimizer and tuning the learning rate until the best model is obtained.

[0034] 3.7 Save the best model: When the loss value reaches a stable state during the training process, the training terminates, and the model with the best performance on the test set is saved.

[0035] Preferably, in S4, the class recognition process is:

[0036] 4.1 Generate class labels: The model predicts the class to which each pixel belongs according to the spectral and spatial features of the input data, combined with the weight parameters learned during the training process. The class label of each pixel represents the corresponding land cover type, including water bodies, vegetation, and bare soil;

[0037] 4.2 Class label post - processing: The generated class labels are subject to post - processing steps, including smoothing filtering and noise optimization;

[0038] 4.3 Generating the ground object class classification map: Based on the post - processed class labels, the model generates a complete ground object class classification map.

[0039] Preferably, conditional random fields and smoothing filtering based on pixel neighborhoods are used to jointly adjust the class labels of adjacent pixels.

[0040] Preferably, the ground object class classification map visualizes the spatial distribution of different ground object classes by annotating the class of each pixel and using color coding.

[0041] Preferably, in S5, a smoothing filtering technique is used to smooth and optimize the classification result. Through filtering operations, discrete error points in the image are eliminated;

[0042] A denoising algorithm is used for denoising processing to detect and correct pixels inconsistent with the ground object classes;

[0043] Spatial consistency optimization is performed by analyzing the class distribution of neighboring pixel points, and conditional random fields and spatial smoothing algorithms are used to optimize and adjust inconsistent regions.

[0044] The present invention discloses a method for rapid identification of ground object classes in the ecological environment based on a convolutional neural network. By optimizing data pre - processing, automatic feature extraction, and efficient result post - processing techniques, the efficiency and classification accuracy of large - scale remote sensing data processing are improved, the computational complexity is reduced, and the generalization ability of the model in different ecological environments is enhanced. Thus, the method of the present invention can achieve rapid and accurate identification of complex ecological environment monitoring tasks and meet the requirements of real - time and large - scale data processing. Brief Description of the Drawings

[0045] Figure 1 It is a flow chart of the ecological environment ground object class recognition technology based on the convolutional neural network of the present invention.

[0046] Figure 2 It is a flow chart of data pre - processing and feature extraction of the present invention.

[0047] Figure 3 It is a flow chart of model training and class recognition of the present invention. Detailed Embodiments

[0048] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0049] The present invention proposes a method for rapid identification of ecological environment ground object categories based on a convolutional neural network. By automatically extracting spectral and spatial features from remote sensing images through the convolutional neural network and combining multi-scale feature extraction and result post-processing techniques, targeted modifications are made to the deficiencies exposed by existing technologies when processing large-scale remote sensing data:

[0050] (1) Improvement for the low identification efficiency of existing technologies

[0051] Existing ground object identification methods (such as rule-based classification or traditional machine learning algorithms) are inefficient when processing large-scale remote sensing data and are difficult to meet the requirements of real-time environmental monitoring. These methods usually rely on manually designed features or manual classification and are difficult to handle complex multi-spectral and high-resolution remote sensing data. The present invention introduces a convolutional neural network (CNN) to achieve automatic feature extraction, significantly improving the classification speed, meeting the requirements of fast and efficient identification, and being applicable to real-time and large-scale ecological monitoring scenarios.

[0052] (2) Improvement for the insufficient classification accuracy

[0053] Remote sensing images contain rich spectral and spatial information. Existing methods cannot make full use of this information for accurate classification. Especially in complex scenarios and fine-grained classification tasks, misclassification or missed classification is likely to occur. The present invention optimizes the structure of the convolutional neural network to deeply mine multi-scale image features, thereby improving the classification accuracy, reducing misclassification and missed classification, and being particularly suitable for processing remote sensing data in complex ecological environments, improving the reliability and accuracy of classification results.

[0054] (3) Improvement for the high complexity of data processing

[0055] With the continuous increase in the resolution and data volume of remote sensing images, existing technologies show problems such as low efficiency, high computational cost, and long processing time when processing multi-band and multi-dimensional remote sensing data. The present invention optimizes the data preprocessing process and combines an efficient convolutional neural network architecture to reduce the computational complexity and improve the data processing efficiency, enabling the system to efficiently process large-scale remote sensing data with limited computing resources and meet the monitoring requirements of complex environments.

[0056] (4) Improvement for the insufficient model generalization ability

[0057] Traditional ground object classification models have weak generalization ability and unstable classification results when facing different scenarios or remote sensing data sources. The present invention enhances the adaptability and robustness of the model by introducing techniques such as data augmentation and regularization, ensuring that it can classify stably and accurately under different ecological environments and data sources, and improving the wide adaptability of the model in practical applications.

[0058] Such as Figure 1As shown in the figure, it is the overall process architecture diagram of the rapid recognition method for ecological environment ground object categories of the present invention. It can be seen from Figure 1 that the overall process is as follows:

[0059] First, the system receives and processes hyperspectral remote sensing images. Through data preprocessing steps, including dimensionality reduction, noise removal, normalization, and data augmentation, the data quality is ensured. Through preprocessing such as dimensionality reduction and denoising, redundant information and noise are effectively removed, key features are retained, and the stability and robustness of the model are enhanced.

[0060] Secondly, based on the convolutional neural network, feature extraction is performed on the preprocessed hyperspectral remote sensing data. Convolution and pooling operations extract multi-scale spectral and spatial features, comprehensively capturing local and global information of different ground object categories, laying a foundation for subsequent classification tasks.

[0061] In the model training stage, the system uses the labeled training set for model learning. By setting the learning rate and hyperparameters, the cross-entropy loss function is used to calculate the difference between the predicted label and the true label, and the network weights are adjusted through backpropagation. After multiple iterations and cross-validations, the loss value is gradually reduced until the model reaches the optimal state.

[0062] After training is completed, the test set is input into the trained model for classification prediction. The model classifies each pixel based on spectral and spatial features, generating a classification result map of ground object categories, showing the spatial distribution of each category.

[0063] In the result post-processing stage, through smoothing processing, noise removal, and spatial consistency optimization, the coherence and accuracy of the classification results are ensured, and finally a high-quality ground object classification map is output.

[0064] Furthermore, a detailed introduction is given to the rapid recognition method for ecological environment ground object categories based on the convolutional neural network proposed in the present invention, including five main steps: data preprocessing, feature extraction, model training, category recognition, and result post-processing. Through the above steps of processing, the accuracy, robustness, and generalization ability of the model in ground object category recognition are effectively improved. Specifically:

[0065] S1. Data preprocessing

[0066] Hyperspectral remote sensing data has characteristics such as high dimensionality and multiple bands, usually containing a large amount of redundant information and noise, which directly affect the efficiency and accuracy of subsequent model training. To improve data quality and enhance the stability and generalization ability of the model, the preprocessing link includes steps such as dimensionality reduction, normalization, noise removal, and data augmentation. As Figure 2 shown, the data preprocessing includes the following processes:

[0067] (1) Dimensionality reduction processing

[0068] Hyperspectral remote sensing images usually contain hundreds of bands, many of which have redundant information, increasing the computational burden and reducing the classification accuracy. To improve efficiency, principal component analysis (PCA) is used to reduce the dimensionality of the data. The covariance matrix between bands is calculated, the main components are extracted, the data dimension is compressed, and the key spectral information is retained, while significantly improving the computational efficiency of the model.

[0069] (2) Normalization

[0070] The numerical ranges of different bands vary greatly. Directly inputting them into the model may lead to unstable training and affect the convergence speed. Through normalization, all band data are scaled to a unified range (such as 0 to 1), avoiding the excessive influence of specific bands on the model, thereby improving the training stability and accelerating the model convergence.

[0071] (3) Denoising

[0072] During the acquisition of hyperspectral remote sensing images, they are often interfered by factors such as sensors and the environment, resulting in noise being mixed into the data. To ensure the quality of the input data, Gaussian filtering or median filtering techniques are used to remove noise, eliminate outliers, and retain important spectral and spatial features to ensure that the model can process clearer data.

[0073] (4) Data augmentation

[0074] To prevent the model from overfitting during training and enhance the generalization ability of the model, data augmentation is performed on the training dataset. Through operations such as random rotation, scaling, and flipping, the diversity of the data is enhanced, and the scale of the dataset is further expanded, thereby improving the model's adaptability to different scenarios and significantly improving the classification accuracy and robustness.

[0075] S2. Feature extraction

[0076] After completing the data preprocessing, in the feature extraction stage, a convolutional neural network is used to efficiently capture the spectral and spatial information in hyperspectral remote sensing data. The present invention adopts a residual network (ResNet) architecture. By automatically extracting multi-scale features, the model's ability to capture complex ground object features is enhanced, effectively solving the problem of gradient disappearance in the training of deep networks, and further improving the classification accuracy and generalization ability of the model. As Figure 2 shown, the feature extraction includes the following process:

[0077] (1) ResNet network architecture

[0078] ResNet solves the vanishing gradient problem in the training of deep networks through the skip connections mechanism. This network consists of multiple residual blocks, and each residual block is composed of multiple convolutional layers and skip connections. Through the skip connections, the input signal can be directly transmitted to deeper layers of the network while convolutional processing is carried out. This design speeds up the network convergence rate and enhances the ability to extract complex and diverse features.

[0079] (2) Convolution operation

[0080] In the convolutional layer, a 3×3 convolutional kernel is used to perform convolution operations on the input data. The convolutional kernel slides in the hyperspectral image, gradually extracting local features such as edges, textures, and shapes. These local features are crucial for the accurate recognition of different land cover classes, ensuring that the classification model has strong feature expression capabilities.

[0081] (3) Pooling operation

[0082] After the convolution operation, a max pooling operation is adopted to reduce the spatial dimension of the feature map while retaining the most representative features. The pooling operation improves the processing speed of the model by reducing the amount of data and effectively prevents overfitting.

[0083] (4) Spectral and spatial feature extraction

[0084] Each pixel of the hyperspectral remote sensing image contains multiple spectral bands, providing rich spectral reflectance information. Through the ResNet network architecture, the model can effectively capture the differences between spectral bands and achieve accurate discrimination of different land cover classes. At the same time, multiple convolutional operations in the network can extract spatial features layer by layer, identify the shape, structure, and spatial distribution of land covers, thereby further enhancing the classification accuracy and the robustness of the model.

[0085] (5) Optimized residual block

[0086] By introducing skip connections, information can be transmitted losslessly in the deep network, ensuring that the model can learn both shallow features and capture deep features. The optimized ResNet structure not only greatly improves the training efficiency of the model but also enhances the integrity and accuracy of feature extraction, laying a solid foundation for subsequent classification tasks.

[0087] S3. Model training

[0088] After feature extraction is completed, in the model training stage, a large amount of labeled data sets are used to train the convolutional neural network to optimize the model parameters so that it can accurately classify hyperspectral remote sensing data. The present invention designs a lightweight CNN model and uses a large-scale labeled hyperspectral remote sensing data set for training. Through parameter optimization and weight adjustment, the model has efficient ground object classification capabilities and reduces the loss value in multiple iterations to ensure its accurate recognition of complex ground object categories. During the entire training process, the network updates the weights through an optimizer to improve the generalization ability of the model and avoid overfitting.

[0089] The model training process is as Figure 3 shown, specifically:

[0090] (1) Dataset division and preprocessing

[0091] The dataset is divided into a training set and a test set. The training set is used for training the network model, and the test set is used to evaluate the generalization ability of the model. To improve the robustness of the model, the training data is processed through data augmentation (such as random rotation, scaling, flipping, etc.) to expand the diversity of the dataset and ensure the stable performance of the model in multiple scenarios.

[0092] (2) Network model initialization

[0093] The initialization of the convolutional neural network is based on the Residual Network (ResNet) architecture. By introducing residual connections, the depth is ensured to be stable when extracting spectral and spatial features. The initial weight parameters are set randomly to provide a good initialization state for the model.

[0094] (3) Loss function calculation

[0095] During training, the cross-entropy loss function is used to evaluate the difference between the predicted labels and the true labels. The loss function measures the classification performance of the current model. The larger the error, the worse the model performance, and thus further optimization is required.

[0096] (4) Error backpropagation and weight update

[0097] The loss value is propagated layer by layer to each layer of the network through backpropagation, and the optimizer updates the weight parameters according to the loss value. The optimizer controls the step size of each weight update by adjusting the learning rate to ensure the gradual convergence of the model.

[0098] (5) Iterative training and parameter adjustment

[0099] The model is continuously optimized in multiple rounds of iterative training, and each round of training is accompanied by the update of the network weights. Hyperparameters such as the learning rate and batch size are dynamically adjusted according to the performance of the model to ensure the efficient convergence of the training process and avoid overfitting problems.

[0100] (6) Model Validation and Performance Optimization

[0101] After each round of training is completed, the generalization ability and classification accuracy of the model are verified through the test set. The model performance is gradually improved through optimizer adjustment and learning rate tuning, and finally an optimal model with good generalization ability is obtained.

[0102] (7) Saving the Optimal Model

[0103] When the loss value reaches a stable state during the training process, the training terminates, and the model with the best performance on the test set is saved for subsequent class recognition tasks. The final model can accurately identify the land cover classes, meeting the requirements of practical applications.

[0104] S4. Class Recognition

[0105] The trained model is applied to new hyperspectral remote sensing data for class recognition. By inputting test data or real-time data, the model classifies each pixel or region based on the features learned previously, generates class labels, and outputs a complete land cover classification map. As Figure 3 shown, the class recognition process includes:

[0106] (1) Generating Class Labels

[0107] Based on the spectral and spatial features of the input data and combined with the weight parameters learned during the training process, the model predicts the class to which each pixel belongs. The class label of each pixel represents its corresponding land cover type, such as water body, vegetation, bare soil, etc. This step ensures efficient and accurate classification of each pixel through the automatic feature extraction of the convolutional neural network.

[0108] (2) Post-processing of Class Labels

[0109] To further improve the accuracy of the recognition results, the generated class labels usually go through a series of post-processing steps, including smoothing filtering and noise optimization, to reduce noise in the recognition process and enhance spatial consistency. Conditional Random Field (CRF) and smoothing filtering based on pixel neighborhood are used to jointly adjust the class labels of adjacent pixels, eliminating isolated misclassifications and making the class distribution smoother and more consistent.

[0110] (3) Generating the Land Cover Classification Map

[0111] Based on the post-processed class labels, the model generates a complete land cover classification map. This classification map visually shows the spatial distribution of different land cover classes in a color-coded manner, providing an intuitive basis for subsequent analysis and interpretation. By labeling the class of each pixel, the generated land cover classification map can clearly show the distribution of each class in different regions and play an important role in fields such as ecological environment monitoring and resource assessment.

[0112] S5, Result Post-Processing

[0113] After generating the preliminary classification results through category recognition, there may be noise or isolated misclassification points. To further improve the accuracy of the classification results and the image quality, the present invention designs a dedicated result post-processing step, mainly including smoothing processing, denoising, and spatial consistency optimization.

[0114] (1) Smoothing Processing

[0115] The initially generated classification map may have discontinuous classification results and isolated misclassification points due to noise or local differences. To solve this problem, the present invention uses smoothing filtering technology to optimize the classification results. Through the filtering operation, discrete error points in the image are eliminated, making the classification results smoother and more coherent, and improving the consistency of vision and analysis. This processing step can effectively reduce the noise in the boundary area and ensure the overall continuity of the classification results.

[0116] (2) Denoising Processing

[0117] In the classification results, some noise points may cause misclassification marks. To eliminate this noise, the present invention uses a denoising algorithm for processing. The denoising algorithm detects and corrects pixels that are inconsistent with the ground object category, eliminating misclassifications. Especially in complex boundary areas and areas where ground objects are similar, the denoising process can effectively improve the classification accuracy and enhance the robustness of the model. The denoising algorithms used include Gaussian filtering, median filtering, or algorithms based on neighborhood consistency, ensuring a high classification accuracy of the processed image.

[0118] (3) Spatial Consistency Optimization

[0119] To ensure the spatial consistency of the classification map, the present invention introduces spatial consistency optimization. In some cases, due to the uncertainty of model prediction, the class labels of neighboring pixels may be different. By analyzing the class distribution of neighboring pixel points, the present invention uses a conditional random field (CRF) and a spatial smoothing algorithm to optimize and adjust inconsistent areas. This step ensures that the classification results are reasonable and consistent in space, eliminates isolated misprediction points, and improves the overall credibility of the classification results.

[0120] In summary, the present invention proposes a fast recognition method for ecological environment ground object categories based on convolutional neural networks, providing an efficient and accurate solution to the problems of low recognition efficiency, insufficient classification accuracy, and difficulty in processing complex multi-spectral data in the prior art. Through data preprocessing, multi-scale feature extraction, optimized model training, accurate category recognition, and post-processing operations of the results, this method can achieve efficient ground object classification in large-scale remote sensing image data and is applicable to multiple application scenarios such as ecological environment monitoring, land use analysis, and environmental change assessment.

[0121] Compared with traditional rule-based classification methods, the present invention makes full use of the advantages of convolutional neural networks in automatic feature extraction and learning, reduces the dependence on manual intervention, and significantly improves the classification accuracy and efficiency. Through means such as data augmentation and regularization, the generalization ability of the model is effectively improved, and it can adapt to various remote sensing data types and complex ground object scenarios.

[0122] In the post-processing stage of the results, the accuracy of the classification map is further optimized. Through smoothing filtering, boundary optimization, and small area merging, misclassifications are effectively reduced, making the spatial distribution of ground object categories more coherent and clear.

[0123] Therefore, the recognition speed and classification accuracy of this method are improved compared with traditional methods, and it is particularly suitable for large-scale and multi-scenario ecological environment monitoring tasks. The present invention not only improves the efficiency and accuracy of ground object recognition but also provides high-quality ecological environment data support for decision-makers, providing a strong technical guarantee for land management, resource planning, and ecological protection.

[0124] The above embodiments are not limitations on the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions, or substitutions made by those skilled in the art within the scope of the technical solution of the present invention also fall within the protection scope of the present invention.

Claims

1. A method for rapid identification of ecological environment features based on convolutional neural network, characterized by: The identification steps include: S1. Data preprocessing: Data preprocessing of hyperspectral remote sensing images; S2. Feature extraction: Feature extraction is performed on the pre-processed hyperspectral remote sensing data based on a convolutional neural network. Multi-scale spectral and spatial features are extracted through convolution and pooling operations to fully capture the local and global information of different types of objects. S3, model training: Use the labeled training set to learn the model, set the learning rate and hyperparameters, use the cross entropy loss function to calculate the difference between the predicted label and the true label, and adjust the network weights through back propagation; After multiple iterations and cross-validation, the loss value is gradually reduced until the model reaches the optimal state; S4. Category recognition: After the training is completed, the test set is input into the trained model for classification prediction; The model classifies each pixel based on spectral and spatial features and generates a classification result map of the ground object category; S5, result post-processing: Post-process the classification result map generated by S4, including smoothing, denoising and spatial consistency optimization, and finally output a high-quality land feature classification map.

2. The method for rapid identification of ecological environment features based on convolutional neural network according to claim 1 is characterized in that: In S1, the preprocessing methods include dimensionality reduction, noise removal, normalization, and data enhancement.

3. The method for rapid identification of ecological environment features based on convolutional neural network according to claim 2 is characterized in that: The principal component analysis method is used to reduce the dimension of the data, calculate the inter-band covariance matrix, extract the main components, compress the data dimension, and retain the key spectral information; Use Gaussian filtering or median filtering techniques to remove noise; All band data are scaled to a uniform range through normalization; Data augmentation processing includes random rotation, scaling, and flipping operations.

4. The method for rapid identification of ecological environment features based on convolutional neural network according to claim 1 is characterized in that: In S2, the feature extraction process is: 2.1 Establish a ResNet network architecture, which consists of multiple residual blocks, each of which consists of multiple convolutional layers and skip connections; 2.2 Convolution operation: In the convolution layer, a 3×3 convolution kernel is used to perform convolution operation on the input data; 2.3 Pooling operation: After the convolution operation, the maximum pooling operation is used to reduce the spatial dimension of the feature map while retaining the most representative features; 2.4 Spectral and spatial feature extraction: Through the ResNet network architecture, the model captures the differences between spectral bands and distinguishes different types of objects; through multi-layer convolution operations, spatial features are extracted layer by layer to identify the shape, structure and spatial distribution of objects; 2.5 Introduce residual blocks into the ResNet network structure.

5. The method for rapid identification of ecological environment features based on convolutional neural network according to claim 1 is characterized in that: In S3, the model training process is as follows: 3.1 Dataset division and preprocessing: Divide into training set and test set, and perform data enhancement on the training set data, including random rotation, scaling, and flipping; 3.2 Network model initialization: The initialization of the convolutional neural network is based on the ResNet network architecture, and residual connections are introduced; the initial weight parameters are set randomly; 3.3 Loss function calculation: The difference between the predicted label and the true label is evaluated based on the cross entropy loss function; 3.4 Error back propagation and weight update: The loss value is transmitted to each layer of the network layer by layer through back propagation, and the optimizer updates the weight parameters according to the loss value; the optimizer controls the step size of each weight update by adjusting the learning rate; 3.5 Multiple rounds of iterative training and dynamic parameter adjustment; 3.6 Model verification and performance optimization: After each round of training, the generalization ability and classification accuracy of the model are verified through the test set; the model performance is gradually improved through optimizer adjustment and learning rate tuning until the best model is obtained. 3.7 Save the best model: When the loss value reaches a stable state during training, the training is terminated and the model with the best performance on the test set is saved.

6. The method for rapid identification of ecological environment features based on convolutional neural network according to claim 1 is characterized in that: in S4, the category identification process is: 4.1 Generate category labels: The model predicts the category of each pixel based on the spectral and spatial characteristics of the input data and the weight parameters learned during the training process. The category label of each pixel represents the corresponding ground feature type, including water, vegetation, and bare soil. 4.2 Category label post-processing: The generated category labels undergo post-processing steps, including smoothing filtering and noise optimization; 4.3 Generate a classification map of land features: Based on the post-processed category labels, the model generates a complete classification map of land features.

7. The method for rapid identification of ecological environment features based on convolutional neural network according to claim 1 is characterized by: Conditional random fields and pixel neighborhood-based smoothing filtering are used to jointly adjust the category labels of adjacent pixels.

8. The method for rapid identification of ecological environment features based on convolutional neural network according to claim 7 is characterized in that: The land object classification map labels the category of each pixel and visualizes the spatial distribution of different land object categories in a color-coded manner.

9. The method for rapid identification of ecological environment features using convolutional neural networks according to claim 1, characterized in that: In S5, smoothing filtering technology is used to smooth and optimize the classification results, and the discrete error points in the image are eliminated through filtering operations; De-noising algorithm is used to perform denoising processing, detect and correct pixels that are inconsistent with the ground object category; The spatial consistency is optimized by analyzing the category distribution of neighboring pixels, and the inconsistent areas are optimized and adjusted using conditional random fields and spatial smoothing algorithms.

Images