Channel cultivated land identification method based on deep convolution and deep belief network
By constructing and training the DCNN-DBN joint network model, combining the characteristics of deep convolution and deep confidence network, the misjudgment and missed detection of farmland recognition under complex channel surface coverage types is solved, and efficient and accurate farmland recognition effect is achieved.
Patent Information
- Application Number
- CN202411769079.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-04
- Publication Date
- 2025-03-28
AI Technical Summary
The prior art is difficult to efficiently and accurately identify cultivated land under complex and changeable channel surface coverage types, and it is prone to misjudgment or missed inspection.
Using a joint network model based on deep convolution and deep confidence network, the channel farmland is recognized by constructing and training the neural network structure of DCNN-DBN semi-supervised learning, combining the spatial feature extraction ability of the deep convolution network and the unsupervised feature self-learning ability of the deep confidence network.
The accuracy and robustness of farmland recognition are significantly improved, and efficient learning and recognition of the characteristics of the trench farmland are achieved, especially when the sample size is small, the recognition accuracy can be maintained.
Smart Images
Figure CN119723365A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of image processing technology, and in particular to a method for identifying ditch cultivated land based on deep convolution and deep belief network. Background Art
[0002] In the field of modern agriculture and land resource management, cultivated land is an important land use type, and its accurate identification is of vital importance to ensure food security, optimize land resource allocation, and implement refined agricultural management. With the rapid development of remote sensing technology, it has become possible to obtain large-scale, high-resolution cultivated land image data, but how to efficiently and accurately extract cultivated land information from these massive data is still a technical problem that needs to be solved. Traditional image processing and pattern recognition methods often find it difficult to balance efficiency and accuracy, especially when faced with complex and changeable gully surface cover types, which are prone to misjudgment or missed detection. Summary of the invention
[0003] The embodiment of the present application provides a method for identifying ditch cultivated land based on deep convolution and deep belief network, so as to solve the problem that misjudgment or missed detection is easy to occur in the prior art when dealing with complex and changeable ditch surface coverage types.
[0004] The embodiment of the present application provides a method for identifying ditch cultivated land based on deep convolution and deep belief network, including:
[0005] Construct a joint network model, which includes a parallel deep convolutional network and a deep belief network. The output ends of the deep convolutional network and the deep belief network are connected to the feature fusion module, and the output end of the feature fusion module is connected to the classifier.
[0006] Acquire remote sensing images to be identified;
[0007] Extract key feature data from remote sensing images to be identified;
[0008] Input the remote sensing image to be identified and the key feature data into the deep convolutional network and the deep belief network respectively to obtain the first feature vector and the second feature vector respectively;
[0009] Inputting the first feature vector and the second feature vector into a feature fusion module, fusing the first feature vector and the second feature vector to obtain a fused feature vector;
[0010] The fused feature vector is input into the classifier, and the fused feature vector is classified to obtain the ditch cultivated land identification result.
[0011] In one implementation method, after constructing the joint network model, the joint network model is trained, and the training method includes: collecting remote sensing images of the ditch terrain area; labeling the land types in the remote sensing images to obtain labeled remote sensing images, and multiple labeled remote sensing images constitute a labeled image data set; dividing the labeled image data set into a training set, a test set, and a validation set; inputting the training set into the joint network model to obtain training recognition results; inputting the training recognition results and the corresponding annotations into a loss function to determine the loss function value; adjusting the hyperparameters of the joint network model according to the loss function value to obtain the trained joint network model; and using the test set and the validation set to test and verify the trained joint network model respectively.
[0012] In one possible implementation, the loss function uses a cross entropy loss function.
[0013] In one possible implementation, when adjusting the hyperparameters of the joint network model, a first hyperparameter is searched in the hyperparameter space of the first network layer to obtain a first hyperparameter, the first hyperparameter is subjected to Bayesian optimization processing and then transmitted to the second network layer, the second network layer searches in the hyperparameter space to obtain a second hyperparameter, the second hyperparameter is combined with the first hyperparameter subjected to Bayesian optimization processing to obtain a second comprehensive hyperparameter, the second comprehensive hyperparameter is subjected to Bayesian optimization processing and then transmitted to the third network layer, until the hyperparameters of all network layers are optimized.
[0014] In one possible implementation, the feature fusion module uses the fused feature vector generated during the training process as the hyperparameter space of each layer of the deep convolutional network. The feature fusion module performs feature matrix decomposition on the fused feature vector generated during the training process to obtain a multi-dimensional feature vector, and uses the multi-dimensional feature vector as the hyperparameter space of each layer of the deep belief network.
[0015] In one possible implementation, the deep convolutional network and the deep belief network simultaneously search for hyperparameters in the corresponding hyperparameter space for parameter optimization.
[0016] In one possible implementation, a random search is used to search for hyperparameters in the hyperparameter space.
[0017] In one possible implementation, the key feature data includes digital elevation model, slope, aspect, and relief index.
[0018] The ditch farmland identification method based on deep convolution and deep belief network in this application has the following advantages:
[0019] DCNN (Deep Convolutional Neural Network) has a strong ability to extract spatial features, and DBN (Deep Belief Nets) can perform unsupervised feature self-learning. Therefore, the deep learning network model DCNN-DBN, which combines DCNN and DBN, can effectively combine supervised classification and unsupervised classification, which is especially suitable for situations like the lack of remote sensing image samples of ditch cultivated land. By constructing the DCNN-DBN semi-supervised learning neural network structure, the unique spectral, texture and spatial structure characteristics of ditch cultivated land can be captured more accurately, significantly improving the accuracy and robustness of cultivated land identification, and realizing efficient learning and identification of ditch cultivated land characteristics. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0021] Figure 1 An architectural diagram of a ditch farmland identification method based on deep convolution and deep belief network provided in an embodiment of the present application.
[0022] Figure 2 An architectural diagram of the parameter optimization process provided in an embodiment of the present application. DETAILED DESCRIPTION
[0023] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0024] Figure 1 The present invention provides an architecture diagram of a method for identifying ditch cultivated land based on deep convolution and deep belief network according to an embodiment of the present invention. The present invention provides a method for identifying ditch cultivated land based on deep convolution and deep belief network, including:
[0025] S100, constructing a joint network model, the joint network model includes a parallel deep convolutional network and a deep belief network, the output ends of the deep convolutional network and the deep belief network are connected to a feature fusion module, and the output end of the feature fusion module is connected to a classifier.
[0026] Exemplarily, after the joint network model is constructed, the joint network model is trained, and the training method includes:
[0027] S101, collecting remote sensing images of the channel terrain area.
[0028] S102, labeling the land type in the remote sensing image to obtain a labeled remote sensing image, and multiple labeled remote sensing images constitute a labeled image data set.
[0029] S103, dividing the labeled image dataset into a training set, a test set, and a validation set in a ratio of 7:2:1.
[0030] S104, input the training set into the joint network model to obtain the training recognition result.
[0031] S105, inputting the training recognition results and the corresponding annotations into the loss function to determine the loss function value.
[0032] S106, adjusting the hyperparameters of the joint network model according to the loss function value to obtain a trained joint network model.
[0033] S107, using the test set and the validation set to test and validate the trained joint network model respectively.
[0034] Furthermore, the loss function adopts the cross entropy loss function. The cross entropy loss function is adopted in this application because, on the one hand, the suitability of using mean square error to discriminate neurons in feature reconstruction experiments is adopted, and on the other hand, the feature extraction and recognition of land type itself belongs to a binary classification problem. In dealing with binary classification problems, the cross entropy loss function can effectively describe the degree of closeness between the true value and the predicted value. The cross entropy loss function is as follows:
[0035]
[0036] a=f(w·x+b)
[0037] Among them, L represents the loss function value, y represents the labeled value of the sample in the test set, a represents the recognition result output by the joint network model, x represents the network, including the input of the deep convolutional neural network and the deep belief network, n represents the total number of samples, f represents the activation function, and w and b represent the parameters of the network. Due to the change in the network structure, the root mean square error and cross entropy superposition function need to be used simultaneously with the classifier based on the Softmax function in the overall calculation process.
[0038] Furthermore, when adjusting the hyperparameters of the joint network model, a first hyperparameter is searched in the hyperparameter space of the first-layer network, the first hyperparameter is subjected to Bayesian optimization processing and then transmitted to the second-layer network, the second-layer network searches in the hyperparameter space to obtain a second hyperparameter, the second hyperparameter is combined with the first hyperparameter subjected to Bayesian optimization processing to obtain a second comprehensive hyperparameter, the second comprehensive hyperparameter is subjected to Bayesian optimization processing and then transmitted to the third-layer network, until the hyperparameters of all layers of the network are optimized.
[0039] Specifically, the feature fusion module uses the fused feature vectors generated during the training process as the hyperparameter space of each layer of the deep convolutional network for back propagation. The feature fusion module performs feature matrix decomposition on the fused feature vectors generated during the training process to obtain multi-dimensional feature vectors, and uses the multi-dimensional feature vectors as the hyperparameter space of each layer of the deep belief network for back propagation. Specifically, a random search method can be used to search for hyperparameters in the hyperparameter space. In addition, after each layer of the network generates hyperparameters, the hyperparameters will be temporarily stored so that after the feature extraction of the current layer of the network is completed, the hyperparameters will be synchronously processed by Bayesian optimization and transmitted to the next layer of the network.
[0040] In an embodiment of the present application, the deep belief network is formed by sequentially connecting multiple layers of restricted Boltzmann machines (RBMs), each of which includes a visible layer and a hidden layer. Figure 1 The light circles in the figure are the nodes in the visible layer, and the dark circles are the nodes in the hidden layer.
[0041] Furthermore, the deep convolutional network and the deep belief network simultaneously search for hyperparameters in the corresponding hyperparameter space to optimize the parameters. This synchronous optimization method can improve the efficiency of model training.
[0042] S110, obtaining a remote sensing image to be identified.
[0043] For example, high-resolution remote sensing images of the area to be identified may be collected by satellite or aircraft.
[0044] S120, extracting key feature data from the remote sensing image to be identified.
[0045] For example, the key feature data include digital elevation model (DEM), slope, aspect and undulation index, etc.
[0046] S130, inputting the remote sensing image to be identified and the key feature data into a deep convolutional network and a deep belief network respectively, to obtain a first feature vector and a second feature vector respectively.
[0047] S140, inputting the first feature vector and the second feature vector into a feature fusion module, fusing the first feature vector and the second feature vector to obtain a fused feature vector.
[0048] For example, the key feature data is represented by a 1×N-dimensional feature vector, which will be input into the deep belief network to obtain N 1×N-dimensional feature vectors. The deep convolutional neural network will output an N×N-dimensional feature vector, so the fused feature vector formed by combining the feature vector output by the deep convolutional neural network and the feature vector output by the deep belief network will be expressed in the form of a feature matrix. By performing feature matrix decomposition on the fused feature vector, N 1×N-dimensional feature vectors can be obtained.
[0049] S150, inputting the fused feature vector into a classifier, classifying the fused feature vector, and obtaining a ditch cultivated land identification result.
[0050] Model Validation
[0051] A remote sensing image with a resolution of 1300×1300 was used, and the corresponding pixels were selected as samples by random sampling. The sample size should be set to be between 20,000 and 50,000. According to actual needs, 20,000, 1,000, and 100 samples were selected in this experiment. In the experiment, the training set and the validation set were divided according to a ratio of 7:3. In order to verify the performance of the proposed method and the current mainstream network model on the data, the overall recognition accuracy (OA) and Kappa value were used as the evaluation of recognition accuracy, and the number of parameters (Parameters), floating-point operations per second (FLOPs) and time (Times) were used as the evaluation of model performance. Since the proposed method is targeted, especially in terms of sample learning and sample quantity, it has good recognition results, so the data is subjected to a sample reduction experiment to verify the influence of the training sample size on the method. According to the current mainstream remote sensing image classification method, the experiment selected the traditional classification method Support Vector Machine (SVM). The deep learning methods mainly include: two models based on deep convolutional neural networks VGGNet (Visual Geometry Group Network) and Unet++, and an unsupervised classification network model DCCM.
[0052] Ditch cultivated land identification results:
[0053] Table 1 Results of 20,000 samples
[0054] method OA (%) Kappa value (%) Parameters(M) FLOPs(G) Times(h) SVM 77.81 79.11 20.23 4.71 0.28 VGGNet 82.31 80.55 59.35 93.77 7.20 Unet++ 85.46 81.19 65.85 279.64 6.11 DCCM 83.99 81.02 37.88 231.22 3.20 DCNN-DBN 92.78 90.77 35.44 54.06 3.70
[0055] Table 2 Results of 1000 samples
[0056] method OA (%) Kappa value (%) Parameters(M) FLOPs(G) Times(h) VGGNet 77.29 72.46 33.37 42.89 4.11 Unet++ 80.12 78.33 40.25 160.72 3.01 DCCM 83.99 81.02 37.88 231.22 3.70 DCNN-DBN 90.69 90.10 20.12 21.06 1.12
[0057] Table 3 Results of 100 samples
[0058] method OA (%) Kappa value (%) Parameters(M) FLOPs(G) Times(h) VGGNet 67.89 66.18 20.11 11.12 0.61 Unet++ 69.10 66.69 19.67 100.11 1.02 DCCM 83.99 81.02 37.88 231.22 3.70 DCNN-DBN 88.59 86.91 20.07 19.09 0.98
[0059] When the sample size is reduced to 100, the classification accuracy and Kappa value of VGGNet and UNet++ have dropped below 70%. The DCNN-DBN network proposed in this application still has good classification and recognition capabilities when the sample size is only 100, and the recognition accuracy can reach 88%, which is only 2% less than when the sample size is 1000. This fully demonstrates the superiority of the DCNN-DBN network model proposed in this application when the sample size is small.
[0060] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.
[0061] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.
Claims
1. A method for identifying ditch cultivated land based on deep convolution and deep belief network, characterized in that: include: Constructing a joint network model, wherein the joint network model includes a parallel deep convolutional network and a deep belief network, wherein the output ends of the deep convolutional network and the deep belief network are both connected to a feature fusion module, and the output end of the feature fusion module is connected to a classifier; Acquire remote sensing images to be identified; Extracting key feature data from the remote sensing image to be identified; Inputting the remote sensing image to be identified and the key feature data into the deep convolutional network and the deep belief network respectively, to obtain a first feature vector and a second feature vector respectively; Inputting the first feature vector and the second feature vector into the feature fusion module, fusing the first feature vector and the second feature vector to obtain a fused feature vector; The fused feature vector is input into the classifier, and the fused feature vector is classified to obtain a ditch cultivated land identification result.
2. The method for identifying ditch cultivated land based on deep convolution and deep belief network according to claim 1 is characterized in that: After the joint network model is constructed, the joint network model is trained, and the training method includes: Collect remote sensing images of the channel terrain area; Labeling the land type in the remote sensing image to obtain the remote sensing image with labels, wherein a plurality of the remote sensing images with labels constitute a label image data set; Splitting the labeled image dataset into a training set, a test set, and a validation set; Inputting the training set into the joint network model to obtain a training recognition result; Input the training recognition results and corresponding annotations into a loss function to determine a loss function value; Adjusting the hyperparameters of the joint network model according to the loss function value to obtain the trained joint network model; The trained joint network model is tested and verified using the test set and the verification set respectively.
3. The method for identifying ditch cultivated land based on deep convolution and deep belief network according to claim 2 is characterized in that: The loss function adopts the cross entropy loss function.
4. The method for identifying ditch cultivated land based on deep convolution and deep belief network according to claim 2 is characterized in that: When adjusting the hyperparameters of the joint network model, a first hyperparameter is searched in the hyperparameter space of the first network layer to obtain a first hyperparameter, the first hyperparameter is subjected to Bayesian optimization processing and then transmitted to the second network layer, the second network layer searches in the hyperparameter space to obtain a second hyperparameter, the second hyperparameter is combined with the first hyperparameter subjected to Bayesian optimization processing to obtain a second comprehensive hyperparameter, the second comprehensive hyperparameter is subjected to Bayesian optimization processing and then transmitted to the third network layer, until the hyperparameters of all network layers are optimized.
5. The method for identifying ditch cultivated land based on deep convolution and deep belief network according to claim 4 is characterized in that: The feature fusion module uses the fused feature vector generated during the training process as the hyperparameter space of each layer of the network in the deep convolutional network. The feature fusion module performs feature matrix decomposition on the fused feature vector generated during the training process to obtain a multidimensional feature vector, and uses the multidimensional feature vector as the hyperparameter space of each layer of the network in the deep belief network.
6. The method for identifying ditch cultivated land based on deep convolution and deep belief network according to claim 5 is characterized in that: The deep convolutional network and the deep belief network simultaneously search for hyperparameters in corresponding hyperparameter spaces to perform parameter optimization.
7. The method for identifying ditch cultivated land based on deep convolution and deep belief network according to claim 4 is characterized in that: A random search method is used to search for hyperparameters in the hyperparameter space.
8. The method for identifying ditch cultivated land based on deep convolution and deep belief network according to claim 1, characterized in that: The key feature data include digital elevation model, slope, aspect and relief index.