Remote sensing image classification and farmland identification method based on deep learning
By adopting deep learning methods, adaptive data augmentation, transfer learning, fine-tuning and mixed precision training in remote sensing image classification and farmland recognition, the problems of long model training time and data differences in different regions in complex farmland block segmentation tasks are solved, and faster and more efficient farmland recognition is achieved.
Patent Information
- Application Number
- CN202411622125.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-14
- Publication Date
- 2025-03-07
AI Technical Summary
In the prior art, the model training time of complex farmland block segmentation tasks may be very long, affecting real-time application, and the farmland data in different regions vary greatly, resulting in limited model application effects.
Remote sensing image classification and farmland recognition methods based on deep learning are adopted, data diversity is improved through RandAugment adaptive data augmentation technology, transfer learning and fine-tuning are used to build a deep feature fusion network, and hybrid precision training technology is used.
The model training time is reduced, the model's generalization ability is improved, so that it can adapt to farmland recognition tasks in different regions and crop types, and reduces training costs.
Smart Images
Figure CN119580088A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of farmland identification, and in particular to a remote sensing image classification and farmland identification method based on deep learning. Background Art
[0002] Remote sensing images are one of the means of obtaining information about the earth's surface through aircraft, satellites, etc., and are widely used in the field of geographic information. Remote sensing image classification and recognition is the process of automatically classifying and identifying these acquired remote sensing images according to categories. It is of great significance for realizing the automated analysis and interpretation of remote sensing images. Through remote sensing image classification and recognition, useful information can be efficiently extracted from massive remote sensing data to provide decision support for urban planning, environmental protection, disaster monitoring and other fields. Farmland identification is an important application in remote sensing image classification.
[0003] After searching, the invention patent with Chinese patent number CN116469006A discloses a method and system for agricultural field segmentation based on deep learning, which belongs to the field of deep learning technology. The method includes the following steps: obtaining agricultural field images from satellite remote sensing farmland data; preprocessing the agricultural field images to obtain a farmland data set, and the farmland data set is divided into a training set and a test set; inputting the farmland data set into a convolutional neural network, and training the convolutional neural network to obtain a convolutional neural network model; inputting the test set in the farmland data set into the convolutional neural network model to obtain an agricultural field segmentation image; wherein the convolutional neural network uses the Mask R-CNN network as a framework, and the convolutional neural network includes a convolutional layer-backbone, a candidate region generation network-RPN, and a network-three branches for classifying candidate regions. This application achieves a faster rate of farmland detection and recognition, and the data is trained on a model with higher accuracy.
[0004] However, in the use of the above patent, for complex farmland segmentation tasks, the training time of the model may be very long, which will not only affect the real-time application of the technical solution, but also increase the workload and time cost of technicians. In addition, there may be large differences in farmland data in different regions, including crop types, growth cycles, soil conditions, etc. These differences may result in limited application effects of the model in different regions. Therefore, a remote sensing image classification and farmland recognition method based on deep learning is proposed. Summary of the invention
[0005] The purpose of this invention is to solve the problem in the prior art that for complex farmland segmentation tasks, the training time of the model may be very long, which will not only affect the real-time application of the technical solution, but also increase the workload and time cost of technicians. In addition, there may be large differences in farmland data in different regions, including crop types, growth cycles, soil conditions, etc. These differences may lead to limited application effects of the model in different regions. A remote sensing image classification and farmland identification method based on deep learning is proposed.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions:
[0007] A remote sensing image classification and farmland identification method based on deep learning includes the following steps:
[0008] S1: Data collection: Obtain remote sensing image data covering the target farmland area from remote sensing satellites, drones and other platforms to ensure that the data resolution, spectral range and time series meet the project requirements;
[0009] S2: Data preprocessing and annotation: Preprocess the remote sensing image data. During the preprocessing, the RandAugment adaptive data enhancement technology is used to increase the diversity of the data by automatically searching for the optimal transformation combination, and the preprocessed data is annotated to generate label data for training and validating the model. The annotation work includes but is not limited to identifying farmland areas, distinguishing different crop types, and annotating other relevant features (such as soil moisture, signs of pests and diseases, etc.);
[0010] S3: Model construction: Construct a convolutional neural network model and use the preprocessed remote sensing image data and annotated data to train the model. During the training process, use the pre-trained deep learning model as a feature extractor and fine-tune the pre-trained model to adapt it to the specific task of farmland recognition in remote sensing images.
[0011] S4: Remote sensing image classification and farmland identification: The remote sensing image to be identified is input into a trained convolutional neural network model, and the convolutional neural network model outputs classification results, including classification information of farmland areas and other land object types.
[0012] The above further includes:
[0013] Further, in S2, in the preprocessing of the remote sensing image data, the RandAugment adaptive data enhancement technology is used to improve the diversity and generalization ability of the remote sensing image data, and the specific steps are:
[0014] Image transformation: Perform random rotation, scaling, flipping and other transformation operations on remote sensing images;
[0015] Color transformation: adjust the color of the remote sensing image, including changes in brightness, contrast and saturation. The color transformation is achieved by adjusting the RGB value of the image;
[0016] Cropping enhancement: Randomly crop a part of the image and splice it with other images or parts of the image. This enhancement method can increase the diversity of training samples. The cropping process is achieved by determining the starting coordinates and size of the cropped area.
[0017] Pixel-level perturbation: Randomly transform the pixel values of remote sensing images, such as adding Gaussian noise, randomly erasing some pixels, etc. These operations improve the model's robustness to noise;
[0018] Automatically search for the optimal strategy: Use RandAugment's search algorithm to automatically find the optimal combination among a large number of data enhancement strategies. This usually involves a training-validation cycle, in which the model is trained on the training set and evaluated on the validation set. Based on the performance on the validation set, the search algorithm adjusts the parameters of the data enhancement strategy and iteratively searches for the optimal combination.
[0019] Furthermore, in S3, the convolutional neural network model is constructed, specifically in the following steps:
[0020] Choose a deep learning framework: Based on project requirements and skill level, choose a suitable deep learning framework, such as TensorFlow, PyTorch, or Keras. These frameworks provide rich APIs and tools to facilitate model building, training, and deployment.
[0021] Constructing a deep feature fusion network: The deep feature fusion network is designed to simultaneously extract global features and local features of remote sensing images to improve the accuracy of farmland recognition;
[0022] Model training: Use preprocessed remote sensing image data and labeled data to train the model;
[0023] Transfer learning and fine-tuning: Using pre-trained deep learning models as feature extractors can significantly reduce data preprocessing costs and training time. At the same time, by fine-tuning the pre-trained models, they can be adapted to the specific task of farmland recognition in remote sensing images.
[0024] Model evaluation and optimization: After model training is completed, the model is evaluated and optimized. Evaluation indicators include accuracy, recall rate, F1 score, etc. Based on the evaluation results, the model is optimized, such as adjusting the network structure, adding data enhancement strategies, using more efficient optimization algorithms, etc.
[0025] Furthermore, in constructing a deep feature fusion network, designing a deep feature fusion network structure design includes the following steps:
[0026] Input layer: receives preprocessed remote sensing image data to ensure that parameters such as image size and color channels are consistent with network input requirements;
[0027] Global feature extraction module: Use a deeper convolutional neural network (such as ResNet, EfficientNet and other pre-trained models) as a global feature extractor. Through transfer learning, use the general feature representation learned by the pre-trained model on a large amount of data, and fine-tune the pre-trained model according to the characteristics of remote sensing images to make it better suited to the task of farmland identification;
[0028] Local feature extraction module: a shallow convolutional neural network is designed to extract local features of remote sensing images. The local feature extraction module includes multiple convolutional layers and pooling layers to gradually extract detailed information in the image.
[0029] Feature fusion layer: The outputs of the global feature extraction module and the local feature extraction module are fused. The fusion can be performed using methods such as concatenation and weighted sum. The fused features will contain global and local information of the remote sensing image, which helps to improve the recognition accuracy of the model.
[0030] Fully connected layer and output layer: After the feature fusion layer, a fully connected layer is added to classify the fused features. The output layer uses the softmax function to output the classification results, that is, the probability distribution of farmland identification.
[0031] Furthermore, in transfer learning and fine-tuning, the specific steps of transfer learning are:
[0032] Select a pre-trained model: Based on the task requirements and data characteristics, select a pre-trained model that performs well in related fields (such as natural image classification);
[0033] Load pre-trained model: Use a deep learning framework (such as TensorFlow or PyTorch) to load the selected pre-trained model, ensuring that the model's weights and architecture are consistent with those during pre-training;
[0034] Adapt the input layer: If the input size of the pre-trained model does not match the size of the remote sensing image data, adjust the size of the input layer;
[0035] Freeze some layers: In order to keep the general feature representation learned by the pre-trained model, freeze some layers of the model (usually the first few layers). The weights of the frozen layers will not be updated during the training process.
[0036] Add custom layers: Based on the pre-trained model, add custom layers suitable for farmland recognition in remote sensing images. The custom layers include additional convolutional layers, fully connected layers, or classification layers.
[0037] Furthermore, in transfer learning and fine-tuning, the specific steps of fine-tuning are:
[0038] Unfreeze some layers: Unfreeze some layers of the pre-trained model (usually layers close to the output) so that the weights of these layers can be updated during training. The number of unfrozen layers depends on the specific requirements of the task and the complexity of the data.
[0039] Set the optimizer and learning rate: Choose an optimizer suitable for the fine-tuning task (such as Adam or SGD) and set a small learning rate to avoid destroying the feature representation learned by the pre-trained model;
[0040] Compile model: compile the model in the deep learning framework, specify the loss function (such as cross entropy loss) and evaluation indicators (such as accuracy);
[0041] Training model: Use preprocessed remote sensing image data and annotated data to train the model. During the training process, monitor the changes in loss and evaluation indicators to adjust the learning rate or perform strategies such as early stopping.
[0042] Evaluate the model: Evaluate the performance of the model on the validation set to ensure that the model is not overfitted. Adjust the model structure or training parameters based on the evaluation results.
[0043] Furthermore, the deep learning framework supports mixed precision training and determines an optimizer suitable for mixed precision training, such as Adam or SGD, which can process FP16 gradients and convert them to FP32 for parameter update when necessary to perform mixed precision training:
[0044] Enable mixed precision training: Before training starts, enable mixed precision training mode, which usually involves setting specific configurations of the deep learning framework or using specific APIs;
[0045] Forward Propagation (FP16): During the forward propagation phase, FP16 precision is used to calculate the output of the model, which helps reduce memory usage and computation time.
[0046] Calculate loss (FP32): When calculating the loss function, convert the model's output and target labels to FP32 precision to ensure the accuracy of loss calculation;
[0047] Backpropagation (FP16 to FP32): In the backpropagation phase, gradients are first calculated using FP16 precision, and then these gradients are converted to FP32 precision as needed for more accurate parameter updates;
[0048] Parameter update (FP32): During the parameter update phase, FP32 precision is used to update the model weights, which helps maintain the accuracy and stability of the model.
[0049] Monitoring and inspection: During training, regularly monitor the model’s loss, accuracy, and other performance metrics to check for any numerical instability or accuracy degradation, and adjust the mixed precision training configuration accordingly.
[0050] Furthermore, in S4, after the convolutional neural network model outputs the classification result, the classification result is post-processed, such as removing noise, filling holes, etc., to improve the accuracy and completeness of the recognition result.
[0051] The present invention has the following beneficial effects:
[0052] 1. In the present invention, the optimal data enhancement strategy is automatically searched through adaptive data enhancement technology, reducing the cost of manually designing data enhancement strategies.
[0053] 2. In the present invention, the deep feature fusion network can simultaneously extract the global features and local features of remote sensing images, enhance the model's ability to recognize complex farmland scenes, and use transfer learning and fine-tuning technology to apply the pre-trained deep learning model to the remote sensing image farmland recognition task, reduce dependence on new data, and improve the generalization ability of the model, so that it can adapt to farmland recognition tasks in different regions, different seasons and different crop types.
[0054] 3. In the present invention, mixed precision training technology is adopted to improve the training speed and reduce the memory usage, thereby reducing the training cost. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 This is a method step diagram for a remote sensing image classification and farmland identification method based on deep learning proposed in the present invention. DETAILED DESCRIPTION
[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0057] See also Figure 1 As shown, the present invention is a remote sensing image classification and farmland identification method based on deep learning, comprising the following steps:
[0058] S1: Data collection: Obtain remote sensing image data covering the target farmland area from remote sensing satellites, drones and other platforms to ensure that the data resolution, spectral range and time series meet the project requirements;
[0059] S2: Data preprocessing and annotation: Preprocess the remote sensing image data. During the preprocessing, the RandAugment adaptive data enhancement technology is used to increase the diversity of the data by automatically searching for the optimal transformation combination, and the preprocessed data is annotated to generate label data for training and validating the model. The annotation work includes but is not limited to identifying farmland areas, distinguishing different crop types, and annotating other relevant features (such as soil moisture, signs of pests and diseases, etc.);
[0060] S3: Model construction: Construct a convolutional neural network model and use the preprocessed remote sensing image data and annotated data to train the model. During the training process, use the pre-trained deep learning model as a feature extractor and fine-tune the pre-trained model to adapt it to the specific task of farmland recognition in remote sensing images.
[0061] S4: Remote sensing image classification and farmland identification: The remote sensing image to be identified is input into a trained convolutional neural network model, and the convolutional neural network model outputs classification results, including classification information of farmland areas and other land object types.
[0062] In one embodiment, for the above S2, in S2, in the preprocessing of the remote sensing image data, the RandAugment adaptive data enhancement technology is used to improve the diversity and generalization ability of the remote sensing image data, and the specific steps are:
[0063] Image transformation: Perform random rotation, scaling, flipping and other transformation operations on remote sensing images;
[0064] Color transformation: adjust the color of the remote sensing image, including changes in brightness, contrast and saturation. The color transformation is achieved by adjusting the RGB value of the image;
[0065] Cropping enhancement: Randomly crop a part of the image and splice it with other images or parts of the image. This enhancement method can increase the diversity of training samples. The cropping process is achieved by determining the starting coordinates and size of the cropped area.
[0066] Pixel-level perturbation: Randomly transform the pixel values of remote sensing images, such as adding Gaussian noise, randomly erasing some pixels, etc. These operations improve the model's robustness to noise;
[0067] Automatically search for the optimal strategy: Use RandAugment's search algorithm to automatically find the optimal combination among a large number of data enhancement strategies. This usually involves a training-validation cycle, in which the model is trained on the training set and evaluated on the validation set. Based on the performance on the validation set, the search algorithm adjusts the parameters of the data enhancement strategy and iteratively searches for the optimal combination.
[0068] In one embodiment, for the above S3, in S3, the convolutional neural network model is constructed, specifically comprising the following steps:
[0069] Choose a deep learning framework: Based on project requirements and skill level, choose a suitable deep learning framework, such as TensorFlow, PyTorch, or Keras. These frameworks provide rich APIs and tools to facilitate model building, training, and deployment.
[0070] Constructing a deep feature fusion network: The deep feature fusion network is designed to simultaneously extract global features and local features of remote sensing images to improve the accuracy of farmland recognition;
[0071] Model training: Use preprocessed remote sensing image data and labeled data to train the model;
[0072] Transfer learning and fine-tuning: Using pre-trained deep learning models as feature extractors can significantly reduce data preprocessing costs and training time. At the same time, by fine-tuning the pre-trained models, they can be adapted to the specific task of farmland recognition in remote sensing images.
[0073] Model evaluation and optimization: After model training is completed, the model is evaluated and optimized. Evaluation indicators include accuracy, recall rate, F1 score, etc. Based on the evaluation results, the model is optimized, such as adjusting the network structure, adding data enhancement strategies, using more efficient optimization algorithms, etc.
[0074] In one embodiment, for the above-mentioned construction of a deep feature fusion network, in the construction of a deep feature fusion network, designing a deep feature fusion network structure design includes the following steps:
[0075] Input layer: receives preprocessed remote sensing image data to ensure that parameters such as image size and color channels are consistent with network input requirements;
[0076] Global feature extraction module: Use a deeper convolutional neural network (such as ResNet, EfficientNet and other pre-trained models) as a global feature extractor. Through transfer learning, use the general feature representation learned by the pre-trained model on a large amount of data, and fine-tune the pre-trained model according to the characteristics of remote sensing images to make it better suited to the task of farmland identification;
[0077] Local feature extraction module: a shallow convolutional neural network is designed to extract local features of remote sensing images. The local feature extraction module includes multiple convolutional layers and pooling layers to gradually extract detailed information in the image.
[0078] Feature fusion layer: The outputs of the global feature extraction module and the local feature extraction module are fused. The fusion can be performed using methods such as concatenation and weighted sum. The fused features will contain global and local information of the remote sensing image, which helps to improve the recognition accuracy of the model.
[0079] Fully connected layer and output layer: After the feature fusion layer, a fully connected layer is added to classify the fused features. The output layer uses the softmax function to output the classification results, that is, the probability distribution of farmland identification.
[0080] In one embodiment, for the above transfer learning and fine-tuning, in the transfer learning and fine-tuning, the specific steps of transfer learning are:
[0081] Select a pre-trained model: Based on the task requirements and data characteristics, select a pre-trained model that performs well in related fields (such as natural image classification);
[0082] For the remote sensing image farmland recognition task, you can choose the ResNet or EfficiencyNet model pre-trained on the ImageNet dataset;
[0083] Load pre-trained model: Use a deep learning framework (such as TensorFlow or PyTorch) to load the selected pre-trained model, ensuring that the model's weights and architecture are consistent with those during pre-training;
[0084] Adapt the input layer: If the input size of the pre-trained model does not match the size of the remote sensing image data, adjust the size of the input layer;
[0085] If the pre-trained model accepts 224x224 input and the remote sensing image is 128x128, the image needs to be upsampled or cropped to fit the input size;
[0086] Freeze some layers: In order to keep the general feature representation learned by the pre-trained model, freeze some layers of the model (usually the first few layers). The weights of the frozen layers will not be updated during the training process.
[0087] Add custom layers: Based on the pre-trained model, add custom layers suitable for farmland recognition in remote sensing images. The custom layers include additional convolutional layers, fully connected layers, or classification layers.
[0088] In one embodiment, for the above transfer learning and fine-tuning, in the transfer learning and fine-tuning, the specific steps of fine-tuning are:
[0089] Unfreeze some layers: Unfreeze some layers of the pre-trained model (usually layers close to the output) so that the weights of these layers can be updated during training. The number of unfrozen layers depends on the specific requirements of the task and the complexity of the data.
[0090] Set the optimizer and learning rate: Choose an optimizer suitable for the fine-tuning task (such as Adam or SGD) and set a small learning rate to avoid destroying the feature representation learned by the pre-trained model;
[0091] Compile model: compile the model in the deep learning framework, specify the loss function (such as cross entropy loss) and evaluation indicators (such as accuracy);
[0092] Training model: Use preprocessed remote sensing image data and annotated data to train the model. During the training process, monitor the changes in loss and evaluation indicators to adjust the learning rate or perform strategies such as early stopping.
[0093] Evaluate the model: Evaluate the performance of the model on the validation set to ensure that the model is not overfitted. Adjust the model structure or training parameters based on the evaluation results.
[0094] Suppose we choose the ResNet50 model pre-trained on ImageNet as the feature extractor and want to apply it to the remote sensing image farmland recognition task. The following are the specific steps of transfer learning and fine-tuning:
[0095] Load the pre-trained model:
[0096] Use PyTorch to load the ResNet50 model and its pre-trained weights;
[0097] Adapting the input layer:
[0098] Resize the input layer to match the remote sensing image data (e.g., 256x256);
[0099] Freeze some layers:
[0100] Freeze the first 4 convolutional blocks of ResNet50 (i.e. conv1_x, conv2_x, conv3_x, conv4_x);
[0101] Add a custom layer:
[0102] After the last convolutional block of ResNet50, add a global average pooling layer, a fully connected layer, and a softmax classification layer;
[0103] To thaw some layers:
[0104] Unfreeze the last convolutional block (conv5_x) and fully connected layers of ResNet50 so that the weights of these layers can be updated during training;
[0105] Set the optimizer and learning rate:
[0106] Select Adam optimizer and set the learning rate to 0.001;
[0107] Compile the model:
[0108] Specify the cross entropy loss function and accuracy evaluation metric;
[0109] Train the model:
[0110] Use preprocessed remote sensing image data and annotated data for training and monitor changes in loss and accuracy;
[0111] Evaluate the model:
[0112] Evaluate the performance of the model on the validation set and make necessary adjustments based on the evaluation results.
[0113] In one embodiment, for the above-mentioned deep learning framework, the deep learning framework supports mixed precision training and determines an optimizer suitable for mixed precision training, such as Adam or SGD, which can process FP16 gradients and convert them to FP32 for parameter update when necessary, and perform mixed precision training:
[0114] Enable mixed precision training: Before training starts, enable mixed precision training mode, which usually involves setting specific configurations of the deep learning framework or using specific APIs;
[0115] Forward Propagation (FP16): During the forward propagation phase, FP16 precision is used to calculate the output of the model, which helps reduce memory usage and computation time.
[0116] In PyTorch, you can use the torch.cuda.amp.autocast() context manager to automatically select the appropriate precision;
[0117] Calculate loss (FP32): When calculating the loss function, convert the model's output and target labels to FP32 precision to ensure the accuracy of loss calculation;
[0118] When defining the loss function, make sure the input is of FP32 type;
[0119] Backpropagation (FP16 to FP32): In the backpropagation phase, gradients are first calculated using FP16 precision, and then these gradients are converted to FP32 precision as needed for more accurate parameter updates;
[0120] In PyTorch, you can use torch.cuda.amp.grad_scaler to automatically handle gradient scaling and updates;
[0121] Parameter update (FP32): During the parameter update phase, FP32 precision is used to update the model weights, which helps maintain the accuracy and stability of the model.
[0122] In the optimizer step, ensure that weight updates are based on FP32 precision gradients;
[0123] Monitoring and inspection: During training, regularly monitor the model’s loss, accuracy, and other performance metrics to check for any numerical instability or accuracy degradation, and adjust the mixed precision training configuration accordingly.
[0124] In one embodiment, for the above S4, in S4, after the convolutional neural network model outputs the classification result, the classification result is post-processed, such as removing noise, filling holes, etc., to improve the accuracy and completeness of the recognition result.
[0125] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A remote sensing image classification and farmland identification method based on deep learning, characterized in that: The following steps are involved: S1: Data collection: Acquire remote sensing image data covering the target farmland area; S2: Data preprocessing and annotation: Preprocess the remote sensing image data. During the preprocessing, the RandAugment adaptive data enhancement technology is used to increase the diversity of the data by automatically searching for the optimal transformation combination, and the preprocessed data is annotated to generate label data for training and validating the model; S3: Model construction: Construct a convolutional neural network model and use the preprocessed remote sensing image data and annotated data to train the model. During the training process, use the pre-trained deep learning model as a feature extractor and fine-tune the pre-trained model to adapt it to the specific task of farmland recognition in remote sensing images. S4: Remote sensing image classification and farmland identification: The remote sensing image to be identified is input into a trained convolutional neural network model, and the convolutional neural network model outputs classification results, including classification information of farmland areas and other land object types.
2. The method for remote sensing image classification and farmland identification based on deep learning according to claim 1, characterized in that: In S2, in the preprocessing of the remote sensing image data, the RandAugment adaptive data enhancement technology is used to improve the diversity and generalization ability of the remote sensing image data. The specific steps are: Image transformation: Perform random transformation operations on remote sensing images; Color transformation: adjust the color of remote sensing images, including changes in brightness, contrast and saturation; Crop augmentation: randomly crop a part of the image and splice it with other images or parts of the image; Pixel-level perturbation: Randomly transform the pixel values of remote sensing images; Automatically search for the optimal strategy: Use RandAugment's search algorithm to automatically find the optimal combination among a large number of data enhancement strategies.
3. The method for remote sensing image classification and farmland identification based on deep learning according to claim 1, characterized in that: In S3, the convolutional neural network model is constructed, specifically in the following steps: Choose a deep learning framework: Choose a deep learning framework; Constructing a deep feature fusion network: The deep feature fusion network is designed to simultaneously extract global features and local features of remote sensing images; Model training: Use preprocessed remote sensing image data and labeled data to train the model; Transfer learning and fine-tuning: Use the pre-trained deep learning model as a feature extractor and fine-tune the pre-trained model to adapt it to the specific task of farmland recognition in remote sensing images; Model evaluation and optimization: After model training is completed, the model is evaluated and optimized, and the model is optimized based on the evaluation results.
4. The method for remote sensing image classification and farmland identification based on deep learning according to claim 3, characterized in that: In building a deep feature fusion network, designing a deep feature fusion network structure design includes the following steps: Input layer: receives preprocessed remote sensing image data; Global feature extraction module: uses a deeper convolutional neural network as a global feature extractor. Through transfer learning, it uses the general feature representation learned by the pre-trained model on a large amount of data to fine-tune the pre-trained model according to the characteristics of remote sensing images. Local feature extraction module: a shallow convolutional neural network is designed to extract local features of remote sensing images. The local feature extraction module includes multiple convolutional layers and pooling layers to gradually extract detailed information in the image. Feature fusion layer: fuses the outputs of the global feature extraction module and the local feature extraction module. The fused features will contain global and local information of the remote sensing image. Fully connected layer and output layer: After the feature fusion layer, a fully connected layer is added to classify the fused features. The output layer uses the softmax function to output the classification results, that is, the probability distribution of farmland identification.
5. The method for remote sensing image classification and farmland identification based on deep learning according to claim 3, characterized in that: In transfer learning and fine-tuning, the specific steps of transfer learning are: Select a pre-trained model: Based on the task requirements and data characteristics, select a pre-trained model that performs well in related fields; Load pre-trained model: Use the deep learning framework to load the selected pre-trained model, ensuring that the model weights and architecture are consistent with those during pre-training; Adapt the input layer: If the input size of the pre-trained model does not match the size of the remote sensing image data, adjust the size of the input layer; Freeze some layers: Freeze some layers of the model. The weights of the frozen layers will not be updated during the training process. Add custom layers: Based on the pre-trained model, add custom layers suitable for farmland recognition in remote sensing images. The custom layers include additional convolutional layers, fully connected layers, or classification layers.
6. The method for remote sensing image classification and farmland identification based on deep learning according to claim 3, characterized in that: In transfer learning and fine-tuning, the specific steps of fine-tuning are: Unfreeze some layers: Unfreeze some layers of the pre-trained model; Set the optimizer and learning rate: Choose an optimizer suitable for the fine-tuning task and set a small learning rate to avoid destroying the feature representation learned by the pre-trained model; Compile model: compile the model in the deep learning framework, specify the loss function and evaluation metrics; Training model: Use preprocessed remote sensing image data and annotated data to train the model. During the training process, monitor the changes in loss and evaluation indicators. Evaluate the model: Evaluate the performance of the model on the validation set to ensure that the model is not overfitting. Adjust the model structure or training parameters based on the evaluation results.
7. The method for remote sensing image classification and farmland identification based on deep learning according to claim 3, characterized in that: The deep learning framework supports mixed precision training and determines an optimizer suitable for mixed precision training. The optimizer can process FP16 gradients and convert them to FP32 for parameter update when necessary to perform mixed precision training: Enable mixed precision training: Enable mixed precision training mode before training starts. Forward propagation (FP16): In the forward propagation stage, FP16 precision is used to calculate the output of the model; Calculate loss (FP32): When calculating the loss function, convert the model's output and target labels to FP32 precision; Back propagation (FP16 to FP32): In the back propagation phase, the gradients are first calculated using FP16 precision, and then these gradients are converted to FP32 precision as needed; Parameter update (FP32): During the parameter update phase, FP32 precision is used to update the model weights; Monitoring and inspection: During training, regularly monitor the model’s loss, accuracy, and other performance metrics to check for any numerical instability or accuracy degradation, and adjust the mixed precision training configuration accordingly.
8. The method for remote sensing image classification and farmland identification based on deep learning according to claim 1, characterized in that: In S4, after the convolutional neural network model outputs the classification result, the classification result is post-processed.
Citation Information
Patent Citations
Deep learning-based agricultural field segmentation method and system
CN116469006A
Cited By
Incremental depth transfer learning method and system for remote sensing real-time detection of crop diseases and insect pests
CN120673270A
An incremental deep transfer learning method and system for real-time remote sensing detection of crop diseases and pests
CN120673270B
Remote sensing image recognition method, system, equipment and product based on quantum migration
CN120853021A
Soil physicochemical property downscaling method and device, computer equipment and medium
CN121072346A
Vertical array target distance estimation system and method based on twin neural network
CN121541142A