Wetland extraction method and system based on neural network

Abstract

The invention discloses a wetland extraction method based on a neural network. The wetland extraction method comprises the following steps: step 1, acquiring data; the acquisition time interval of the optical remote sensing image data and the acquisition time interval of the radar remote sensing image data have an overlapping part, the overlapping duration is between 30 days and 90 days, and the overlapping duration is comprehensively determined according to the wetland vegetation growth cycle and the hydrological change cycle; the multi-source feature information of the wetland in the key period is reflected by the data; preprocessing operation is carried out on collected original remote sensing image data, and the error rate of radiation correction is strictly controlled within + / -3%. Compared with a traditional artificial remote sensing image interpretation wetland extraction mode, the neural network-based wetland extraction method has the advantages that the processing efficiency is greatly improved, large-scale multi-source remote sensing image data can be quickly processed, and the time and labor cost required by wetland information extraction is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of wetland extraction, and in particular to a wetland extraction method and system based on neural network. Background Art

[0002] Wetlands occupy an important position in the ecosystem, and their monitoring and protection are of vital importance. The traditional method of extracting wetland information by manually interpreting remote sensing images is inefficient and difficult to cope with massive remote sensing data, and can no longer meet current needs. Although the early pixel-based machine learning classification method has a certain efficiency improvement, it is limited to single-source data and ignores spatial information, resulting in defects in the classification results. After the rise of deep learning, neural networks have shown potential in the field of wetland extraction. It can automatically learn features, integrate multi-source remote sensing data (such as optical and radar image features), and use convolution and pooling operations to extract multi-scale features and spatial information, reduce the impact of noise, and adapt to different wetland characteristics through training to improve generalization ability. However, the current neural network-based Wetland extraction via the network still faces challenges. When acquiring data, the time synchronization and quality consistency of multi-source images are difficult to ensure, which will interfere with the accurate extraction of wetland features and model training. The feature extraction and fusion links, the determination of suitable features and effective fusion strategies still need to be further explored. In model construction and training, it is necessary to balance accuracy and efficiency, carefully optimize the structure and parameters, prevent overfitting or underfitting, and ensure rapid convergence and good generalization. In terms of accuracy assessment, traditional indicators are not comprehensive enough, and it is urgent to build a more complete system to measure the performance of the model in diverse wetlands and geographical environments. Therefore, the development of efficient, accurate and adaptable neural network-based wetland extraction methods and systems is of great significance and urgent need for wetland monitoring and management and related scientific research. Summary of the invention

[0003] To achieve the above objectives, the present invention is implemented through the following technical solutions: A wetland extraction method based on neural network, comprising the following steps: Step 1: Get data; Collect multi-source remote sensing image data, where the acquisition time interval of optical remote sensing image data and radar remote sensing image data overlaps, and the overlapping period is between 30 days and 90 days. This overlapping period is determined based on the wetland vegetation growth cycle and hydrological change cycle to ensure that the data reflects the multi-source characteristic information of the wetland in the critical period; Preprocessing operations are performed on the collected raw remote sensing image data. The error rate of radiation correction is strictly controlled within ±3%, the geometric correction registration accuracy is within 0.5 pixels, and the reflectivity error after atmospheric correction is less than ±2%, thus laying the foundation for subsequent accurate feature extraction. Step 2: Extract and fuse features; Extract optical features from preprocessed multi-source remote sensing image data, including spectral features, texture features, and vegetation index features; extract radar features, including backscatter coefficient features and polarization features; A feature fusion method based on principal component analysis (PCA) is used, and the threshold of the principal component contribution rate is set to above 85%, so as to effectively fuse optical and radar features and enhance the representativeness and discrimination of feature information. Step 3: Build and train the neural network model; A wetland extraction neural network model was constructed. The number of nodes in the input layer of the model was determined according to the fused feature dimension. The hidden layer was set to 3 layers. The number of neurons in each layer decreased from the input layer to the output layer in a ratio of 40% to 20%. The output layer was set to 2 nodes, corresponding to the wetland category and the non-wetland category respectively. The constructed neural network model was trained using accurately labeled wetland sample data. The training adopted the stochastic gradient descent optimization algorithm. The initial value of the learning rate was set to 0.01, and it was reduced by 10% every 10 training cycles according to the learning rate decay strategy. The number of training iterations was 100 times. During the training process, the k-fold cross-validation method was used to divide the sample data into training set, validation set and test set at a ratio of 60%, 20% and 20%. Cross-validation was used to ensure the generalization ability and stability of the model. Step 4: Wetland extraction and accuracy assessment; After the multi-source remote sensing image data of the wetland to be extracted is processed through the above-mentioned feature extraction and fusion steps, it is input into the well-trained neural network model to perform the wetland extraction operation and obtain the wetland extraction result; The confusion matrix is ​​used to conduct accuracy evaluation on the wetland extraction results. The overall accuracy OA is required to be no less than 90%, and the user accuracy UA and producer accuracy PA of the wetland category must reach more than 85%. If the accuracy requirements are not met, the parameters of the neural network model are automatically adjusted or the training sample data is expanded, and then the training and extraction process is restarted until the accuracy standards are met.

[0004] Preferably, there is an overlap in the acquisition time interval of the optical remote sensing image data and the radar remote sensing image data in step one, and the overlapping time is between 30 days and 90 days. The error rate of radiation correction during preprocessing is within ±3%, the geometric correction alignment accuracy is within 0.5 pixels, and the reflectivity error after atmospheric correction is less than ±2%.

[0005] Preferably, the optical features in step 2 include spectral features, texture features and vegetation index features, and the radar features include backscatter coefficient features and polarization features.

[0006] Preferably, the number of input layer nodes of the neural network model in step three is determined according to the fused feature dimension, the hidden layer is set to 3 layers, the number of neurons in each layer decreases from the input layer to the output layer at a ratio of 40% to 20%, and the output layer is set to 2 nodes, corresponding to the wetland category and the non-wetland category respectively.

[0007] Preferably, the specific steps of using the confusion matrix to conduct accuracy evaluation on wetland extraction results are as follows: Determine the dimension of the confusion matrix. Since this is a two-classification problem for wetland extraction (wetland category and non-wetland category), the confusion matrix is ​​a 2×2 matrix. The rows represent the actual categories, which are wetland category and non-wetland category respectively; the columns represent the predicted categories, which are also wetland category and non-wetland category; Traverse the entire wetland to extract the result image and the corresponding reference verification image, which is the real wetland distribution image determined by field survey, high-precision mapping or other reliable means; For each pixel in the image, perform the following operations: If it is judged as a wetland category in the extraction result and is also a wetland category in the reference verification image, the count of the first row and first column in the confusion matrix is ​​increased by 1; If it is judged as a wetland category in the extraction result, but is a non-wetland category in the reference verification image, the count of the first column of the second row in the confusion matrix is ​​increased by 1; If it is judged as a non-wetland category in the extraction result, but a wetland category in the reference verification image, the count of the first row and second column in the confusion matrix is ​​increased by 1; If it is judged as a non-wetland category in the extraction result and is also a non-wetland category in the reference verification image, the count of the second row and second column in the confusion matrix is ​​increased by 1; Calculate the accuracy evaluation index based on the statistical quantities in the confusion matrix; The accuracy assessment result is obtained by comparing the calculated overall accuracy, user accuracy and producer accuracy with the set thresholds.

[0008] A wetland extraction system based on neural network, comprising: The data acquisition module is used to acquire multi-source remote sensing image data. It has the function of receiving and storing optical remote sensing image data source and radar remote sensing image data source data, and can perform preliminary screening of the original remote sensing image data according to the image cloud ratio and spatial resolution to ensure data quality and availability. The feature extraction and fusion module is used to extract optical features and radar features from the multi-source remote sensing image data preprocessed by the data acquisition module, and fuse the two; Neural network model building and training module, used to build a wetland extraction neural network model and train it using labeled wetland sample data; The wetland extraction and accuracy assessment module is used to process the multi-source remote sensing image data of the wetland to be extracted through the feature extraction and fusion module, and then input the data into the neural network model trained by the neural network model construction and training module for wetland extraction. The confusion matrix is ​​used to evaluate the accuracy of the wetland extraction results. The overall accuracy OA must be above 90%, and the user accuracy UA and producer accuracy PA of the wetland category must be above 85%. If the standards are not met, the neural network model construction and training module is triggered to adjust the model parameters or increase the training sample data and then retrain and extract until the accuracy is qualified.

[0009] Preferably, a preprocessing unit is provided inside the data acquisition module, and the radiation correction subunit in the unit can accurately control the radiation error within ±3%, the alignment accuracy of the geometric correction subunit is within 0.5 pixels, and the atmospheric correction subunit can make the corrected reflectivity error less than ±2%.

[0010] Preferably, the feature extraction and fusion module also includes an optical feature extraction submodule, a radar feature extraction submodule and a feature fusion submodule, wherein the optical feature extraction submodule can extract spectral features, texture features and vegetation index features; the radar feature extraction submodule can extract backscattering coefficient features and polarization features; the feature fusion submodule adopts a feature fusion method based on principal component analysis PCA, and the principal component contribution rate threshold is set to above 85%.

[0011] Preferably, the training unit in the neural network model construction and training module adopts a stochastic gradient descent optimization algorithm, with an initial learning rate of 0.01, and a strategy adjustment of reducing it by 10% every 10 training cycles. The training is iterated 100 times, and k-fold cross validation is used in training. The sample data is divided into a training set, a validation set, and a test set in a ratio of 60%, 20%, and 20%.

[0012] The present invention provides a wetland extraction method and system based on neural network, which has the following beneficial effects: Compared with the traditional wetland extraction method of manually interpreting remote sensing images, the neural network-based wetland extraction method and system greatly improves the processing efficiency, can quickly process large-scale multi-source remote sensing image data, reduces the time and labor cost required for wetland information extraction, adapts to the timeliness requirements of modern wetland monitoring, effectively integrates optical remote sensing image data and radar remote sensing image data, fully mines and integrates the spectral features, texture features, vegetation index features, backscatter coefficient features and polarization features in multi-source data. This multi-source data fusion method makes the wetland extraction results more comprehensive and accurate, avoids misjudgment or missed judgment caused by insufficient single-source data information, and significantly improves the accuracy of wetland extraction. The neural network has a strong automatic feature learning ability and can directly learn complex and representative wetland feature patterns from the original data without relying on a large number of manually designed features, which not only reduces the artificial feature engineering The deep convolutional neural network can not only reduce the burden of image classification, but also capture some subtle but key wetland features that are difficult to find through traditional methods, further improving the extraction accuracy and adaptability of the model. With the help of convolution and pooling operations in deep convolutional neural networks, the spatial context information of the image can be extracted to a certain extent, overcoming the "salt and pepper noise" problem caused by the traditional pixel classification method ignoring spatial information, making the wetland extraction results more coherent and smooth in space, more in line with the actual geographical distribution characteristics of the wetland, and conducive to the subsequent spatial pattern analysis and ecological function evaluation of the wetland. The neural network model is trained by a large amount of labeled wetland sample data, and cross-validation and other technical means are used. The model can learn the commonalities and differences of wetlands in different regions and types, so that it has good generalization ability and can be applied to wetland extraction tasks in a variety of geographical environments and wetland types, providing unified and reliable technical support for large-scale and multi-scenario wetland monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present invention are described clearly and completely below. 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.

[0015] See also Figure 1 The present invention provides a technical solution: a wetland extraction system based on a neural network, characterized in that it includes: The data acquisition module is used to acquire multi-source remote sensing image data. It has the function of receiving and storing optical remote sensing image data sources and radar remote sensing image data sources, and can preliminarily screen the original remote sensing image data according to the image cloud ratio and spatial resolution to ensure data quality and availability. A pre-processing unit is set up inside. The radiation correction subunit in this unit can accurately control the radiation error within ±3%, the registration accuracy of the geometric correction subunit is within 0.5 pixels, and the atmospheric correction subunit can make the corrected reflectivity error less than ±2%; The feature extraction and fusion module is used to extract optical features and radar features from the multi-source remote sensing image data preprocessed by the data acquisition module, and fuse the two. The optical feature extraction submodule can extract spectral features, texture features, and vegetation index features; the radar feature extraction submodule can extract backscattering coefficient features and polarization features; the feature fusion submodule adopts a feature fusion method based on principal component analysis PCA, and the principal component contribution rate threshold is set to be above 85%; The neural network model construction and training module is used to construct a wetland extraction neural network model and train it using labeled wetland sample data. The training unit uses a stochastic gradient descent optimization algorithm, with an initial learning rate of 0.01, and a strategy adjustment of decreasing by 10% every 10 training cycles. The training is iterated 100 times, and k-fold cross validation (k=5) is used in the training. The sample data is divided into a training set, a validation set, and a test set at a ratio of 60%, 20%, and 20%; The wetland extraction and accuracy assessment module is used to process the multi-source remote sensing image data of the wetland to be extracted through the feature extraction and fusion module, and then input the data into the neural network model trained by the neural network model construction and training module for wetland extraction. The confusion matrix is ​​used to evaluate the accuracy of the wetland extraction results. The overall accuracy OA must be above 90%, and the user accuracy UA and producer accuracy PA of the wetland category must be above 85%. If the standards are not met, the neural network model construction and training module is triggered to adjust the model parameters or increase the training sample data and then retrain and extract until the accuracy is qualified.

[0016] A wetland extraction method based on neural network comprises the following steps: Step 1: Get data; Collect multi-source remote sensing image data, where the acquisition time interval of optical remote sensing image data and radar remote sensing image data overlaps, and the overlapping period is between 30 days and 90 days. This overlapping period is determined based on the wetland vegetation growth cycle and hydrological change cycle to ensure that the data reflects the multi-source characteristic information of the wetland in the critical period; Preprocessing operations are performed on the collected raw remote sensing image data. The error rate of radiation correction is strictly controlled within ±3%, the geometric correction registration accuracy is within 0.5 pixels, and the reflectivity error after atmospheric correction is less than ±2%, thus laying the foundation for subsequent accurate feature extraction. Step 2: Extract and fuse features; Extract optical features from preprocessed multi-source remote sensing image data, including spectral features, texture features, and vegetation index features, and extract radar features, including backscatter coefficient features and polarization features; A feature fusion method based on principal component analysis (PCA) is used, and the threshold of the principal component contribution rate is set to above 85%, so as to effectively fuse optical and radar features and enhance the representativeness and discrimination of feature information. Step 3: Build and train the neural network model; A wetland extraction neural network model was constructed. The number of nodes in the input layer of the model was determined according to the fused feature dimension. The hidden layer was set to 3 layers. The number of neurons in each layer decreased from the input layer to the output layer in a ratio of 40% to 20%. The output layer was set to 2 nodes, corresponding to the wetland category and the non-wetland category respectively. The constructed neural network model was trained using accurately labeled wetland sample data. The training adopted the stochastic gradient descent optimization algorithm. The initial value of the learning rate was set to 0.01, and it was reduced by 10% every 10 training cycles according to the learning rate decay strategy. The number of training iterations was 100 times. During the training process, the k-fold cross-validation method was used to divide the sample data into training set, validation set and test set at a ratio of 60%, 20% and 20%. Cross-validation was used to ensure the generalization ability and stability of the model. Step 4: Wetland extraction and accuracy assessment; After the multi-source remote sensing image data of the wetland to be extracted is processed through the above-mentioned feature extraction and fusion steps, it is input into the well-trained neural network model to perform the wetland extraction operation and obtain the wetland extraction result; Use the confusion matrix to evaluate the accuracy of wetland extraction results. The overall accuracy OA is required to be no less than 90%, and the user accuracy UA and producer accuracy PA of the wetland category must reach more than 85%. If the accuracy requirements are not met, automatically adjust the parameters of the neural network model or expand the training sample data, and then restart the training and extraction process until the accuracy standards are met; The specific steps are as follows: The specific steps for using the confusion matrix to conduct accuracy evaluation of wetland extraction results are as follows: Determine the dimension of the confusion matrix. Since this is a binary classification problem for wetland extraction, the confusion matrix is ​​a 2×2 matrix. The rows represent the actual categories, which are wetland categories and non-wetland categories, respectively, and the columns represent the predicted categories, which are also wetland categories and non-wetland categories. Traverse the entire wetland to extract the result image and the corresponding reference verification image, which is the real wetland distribution image determined by field survey, high-precision mapping or other reliable means; For each pixel in the image, perform the following operations: If it is judged as a wetland category in the extraction result and is also a wetland category in the reference verification image, the count of the first row and first column in the confusion matrix is ​​increased by 1; If it is judged as a wetland category in the extraction result, but is a non-wetland category in the reference verification image, the count of the first column of the second row in the confusion matrix is ​​increased by 1; If it is judged as a non-wetland category in the extraction result, but a wetland category in the reference verification image, the count of the first row and second column in the confusion matrix is ​​increased by 1; If it is judged as a non-wetland category in the extraction result and is also a non-wetland category in the reference verification image, the count of the second row and second column in the confusion matrix is ​​increased by 1; Calculate the accuracy evaluation index based on the statistical quantities in the confusion matrix; The accuracy assessment result is obtained by comparing the calculated overall accuracy, user accuracy and producer accuracy with the set thresholds.

[0017] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field and related fields without creative work should fall within the scope of protection of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention are implemented according to the conventional means in the field unless otherwise specified and limited.

Claims

1. A wetland extraction method based on neural network, characterized in that: The following steps are involved: Step 1: Collect multi-source remote sensing image data and perform preprocessing operations on the collected original remote sensing image data; Step 2: Extract optical features and radar features from the preprocessed multi-source remote sensing image data, and fuse the optical and radar features using a feature fusion method based on principal component analysis (PCA); Step 3: Construct a wetland extraction neural network model and use the accurately labeled wetland sample data to train the constructed neural network model; Step 4: After the multi-source remote sensing image data of the wetland to be extracted is processed by the above-mentioned feature extraction and fusion steps, it is input into the well-trained neural network model to perform the wetland extraction operation, obtain the wetland extraction results, and use the confusion matrix to conduct accuracy evaluation on the wetland extraction results; The multi-source remote sensing image data in step 1 includes radiation correction data, geometric correction data and atmospheric correction data, the error rate of radiation correction is within ±3%, the geometric correction registration accuracy is within 0.5 pixels, and the reflectivity error after atmospheric correction is less than ±2%; The optical features in step 2 include spectral features, texture features and vegetation index features. Radar characteristics include backscatter coefficient characteristics and polarization characteristics.

2. A wetland extraction method based on neural network according to claim 1, characterized in that: The number of input layer nodes of the neural network model in step three is determined according to the fused feature dimension, the hidden layer is set to 3 layers, the number of neurons in each layer decreases from the input layer to the output layer at a ratio of 40% to 20%, and the output layer is set to 2 nodes, corresponding to the wetland category and the non-wetland category respectively.

3. A wetland extraction method based on neural network according to claim 1, characterized in that: The specific steps for using the confusion matrix to conduct accuracy evaluation of wetland extraction results are as follows: Determine the dimensions of the confusion matrix; Traverse the entire wetland to extract the result image and the corresponding reference verification image, which is the real wetland distribution image determined by field survey, high-precision mapping or other reliable means; For each pixel in the image, perform the following operations: If it is judged as a wetland category in the extraction result and is also a wetland category in the reference verification image, the count of the first row and first column in the confusion matrix is ​​increased by 1; If it is judged as a wetland category in the extraction result, but is a non-wetland category in the reference verification image, the count of the first column of the second row in the confusion matrix is ​​increased by 1; If it is judged as a non-wetland category in the extraction result, but a wetland category in the reference verification image, the count of the first row and second column in the confusion matrix is ​​increased by 1; If it is judged as a non-wetland category in the extraction result and is also a non-wetland category in the reference verification image, the count of the second row and second column in the confusion matrix is ​​increased by 1; Calculate the accuracy evaluation index based on the statistical quantities in the confusion matrix; The accuracy assessment result is obtained by comparing the calculated overall accuracy, user accuracy and producer accuracy with the set thresholds.

4. A wetland extraction system based on neural network, characterized in that: include: Data acquisition module, used to acquire multi-source remote sensing image data; The feature extraction and fusion module is used to extract optical features and radar features from the multi-source remote sensing image data preprocessed by the data acquisition module, and fuse the two; Neural network model building and training module, used to build a wetland extraction neural network model and train it using labeled wetland sample data; The wetland extraction and accuracy assessment module is used to process the multi-source remote sensing image data of the wetland to be extracted through the feature extraction and fusion module, input the neural network model trained by the neural network model construction and training module to extract the wetland, and use the confusion matrix to evaluate the accuracy of the wetland extraction results; The data acquisition module is internally provided with a preprocessing unit, in which the radiation correction subunit can accurately control the radiation error within ±3%, the registration accuracy of the geometric correction subunit is within 0.5 pixels, and the atmospheric correction subunit can make the reflectivity error after correction less than ±2%; The feature extraction and fusion module also includes an optical feature extraction submodule, a radar feature extraction submodule and a feature fusion submodule.

5. A neural network-based wetland extraction system according to claim 4, characterized in that: The training unit in the neural network model construction and training module adopts a stochastic gradient descent optimization algorithm, with an initial learning rate of 0.01, and a strategy adjustment of reducing the learning rate by 10% every 10 training cycles, and 100 training iterations.

Images