Tree species classification method and system based on DSC-DC convolutional neural network

Abstract

The application discloses a tree species classification method and system based on a DSC-DC convolutional neural network, and the method comprises the following steps: collecting hyperspectral images of a target area, and performing pretreatment on the hyperspectral images; a DSC-DC classification model is established, which comprises one input layer, seven convolutional layers, two pooling layers, one flattening layer, two fully connected layers, one random inactivation layer and one output layer, wherein the first layer, the third layer, the fifth layer and the eighth layer are separable convolutional layers, the second layer, the fourth layer and the seventh layer are deep convolutional layers, the sixth layer and the ninth layer are pooling layers; the pretreated hyperspectral images are input into the established DSC-DC classification model, and the output tree species classification result is obtained. Through the combination of deep separable convolution and dilated convolution, the application can improve the calculation efficiency while improving the classification accuracy under the condition of multiple tree species and small samples, and provides a powerful solution for the practical application in the field of tree species classification.

Description

Technical Field

[0001] This invention relates to the field of forest tree species classification technology, and in particular to a tree species classification method and system based on DSC-DC convolutional neural network. Background Technology

[0002] Forests, as vital terrestrial ecosystems, provide essential services and resources for human development and well-being. In the context of climate change and its potential impacts on forest ecosystems, it is crucial to recognize that climate change can alter tree species distribution and suitable growing areas, thereby affecting forest composition and distribution. Timely and accurate tree species classification and mapping help forest managers understand forest resource distribution, predict vegetation changes, optimize management plans, and develop sustainable development strategies.

[0003] The rich and continuous bands of hyperspectral data can capture subtle spectral differences related to biogeochemical characteristics, effectively overcoming the phenomenon of "different species with the same spectrum" and "different species with the same spectrum," thus better distinguishing different forest vegetation types, especially under complex forest stand conditions. Therefore, better learning and utilizing these subtle spectral differences is crucial for improving the accuracy of tree species classification.

[0004] To date, machine learning and deep learning methods have been widely used for tree species classification using hyperspectral data. Various machine learning algorithms, such as Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Artificial Neural Networks (ANN), have been applied at different scales and achieved good classification results. However, it is worth noting that machine learning methods rely heavily on human expertise. Therefore, these methods are not only time-consuming and labor-intensive but also neglect consideration of the overall structure and spectral information. Furthermore, the knowledge level of the personnel directly affects the quality of extracted features. Therefore, in the context of large-scale remote sensing data with complex data structures, the automatic extraction of joint spatial-spectral information features becomes a key challenge.

[0005] Compared to machine learning, deep learning offers superior end-to-end learning capabilities, automatically extracting high-level features by effectively integrating spatial-spectral information. In tree classification, network depth is a key factor influencing feature extraction and image processing; therefore, many researchers aim to improve classification accuracy by constructing deeper and more complex network architectures. However, this can increase computational burden and impact the network's real-time performance.

[0006] Therefore, how to provide a tree species classification model and method with high classification accuracy and more efficient training and prediction time for small-sample, multi-tree-species forests is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] The main objective of this invention is to provide a tree species classification method and system based on DSC-DC convolutional neural network, aiming to solve the technical problems of high complexity and low accuracy in current tree species classification models.

[0008] To achieve the above objectives, this invention provides a tree species classification method based on a DSC-DC convolutional neural network, the method comprising the following steps:

[0009] S1: Acquire hyperspectral images of the target area and preprocess the hyperspectral images;

[0010] S2: Establish a DSC-DC classification model consisting of 1 input layer, 7 convolutional layers, 2 pooling layers, 1 flattening layer, 2 fully connected layers, 1 random deactivation layer, and 1 output layer;

[0011] S3: Input the preprocessed hyperspectral image into the established DSC-DC classification model to obtain the output tree species classification results;

[0012] The DSC-DC classification model includes: seven convolutional layers constructed by separable convolutional layers (layers 1, 3, 5, and 8) and depthwise convolutional layers (layers 2, 4, and 7); and two pooling layers constructed by pooling layers (layers 6 and 9).

[0013] Among them, the convolutional layers of the 3rd and 7th layers use dilated convolution with dilation rates of 2 and 4, respectively.

[0014] Optionally, in step S1, acquiring hyperspectral images of the target area and performing preprocessing steps on the hyperspectral images specifically includes:

[0015] S101: Use a canopy height model to digitize individual tree canopies from hyperspectral images and plot data of the target area so that pixels within the canopy can receive species tags;

[0016] S102: In high-resolution aerial images, create circular labels centered on the center point of each tree and label polygons with the species category, so that each sample label contains only pure pixels of the canopy and does not include the background, soil or other canopy elements.

[0017] Optionally, the canopy height model is obtained by processing lidar data of the target area, and the lidar data and the hyperspectral image are configured to be obtained synchronously in the target area.

[0018] Optionally, step S1, which involves acquiring hyperspectral images of the target area and preprocessing the hyperspectral images, further includes: performing dimensionality reduction processing on the cubic data of the acquired hyperspectral images using principal component analysis.

[0019] Optionally, in step S2, the kernel size of the 7 convolutional layers is set to 3×3, and the number of kernels in the next layer is twice that of the previous layer.

[0020] Optionally, in step S2, the flattening layer is configured to flatten the downsampled feature map into a one-dimensional vector.

[0021] Optionally, in step S2, a linear rectified function is used as the activation function for all convolutional and fully connected layers to perform a nonlinear transformation.

[0022] Optionally, in step S2, the random deactivation layer is configured between two fully connected layers.

[0023] Optionally, in step S2, the random deactivation layer is configured to randomly deactivate with a probability of 0.4 to reduce overfitting.

[0024] Furthermore, to achieve the above objectives, the present invention also provides a tree species classification system based on a DSC-DC convolutional neural network, the system comprising:

[0025] The acquisition module is configured to acquire hyperspectral images of the target area and preprocess the hyperspectral images.

[0026] The model building module is configured to build a DSC-DC classification model that includes one input layer, seven convolutional layers, two pooling layers, one flattening layer, two fully connected layers, one random deactivation layer, and one output layer.

[0027] A tree species classification module is configured to input preprocessed hyperspectral images into an established DSC-DC classification model to obtain the output tree species classification results.

[0028] The DSC-DC classification model includes: seven convolutional layers constructed by separable convolutional layers (layers 1, 3, 5, and 8) and depthwise convolutional layers (layers 2, 4, and 7); and two pooling layers constructed by pooling layers (layers 6 and 9).

[0029] Among them, the convolutional layers of the 3rd and 7th layers use dilated convolution with dilation rates of 2 and 4, respectively.

[0030] The beneficial effects of this invention are as follows: It proposes a tree species classification method and system based on a DSC-DC convolutional neural network. The method includes: S1: acquiring hyperspectral images of the target region and preprocessing the hyperspectral images; S2: establishing a DSC-DC classification model comprising one input layer, seven convolutional layers, two pooling layers, one flattening layer, two fully connected layers, one random deactivation layer, and one output layer; S3: inputting the preprocessed hyperspectral images into the established DSC-DC classification model to obtain the output tree species classification results; wherein, the DSC-DC classification model includes: seven convolutional layers jointly constructed by separable convolutional layers (layers 1, 3, 5, and 8) and depthwise convolutional layers (layers 2, 4, and 7); and two pooling layers jointly constructed by pooling layers (layers 6 and 9); wherein the convolutional layers of layers 3 and 7 employ dilated convolution with dilation rates of 2 and 4, respectively. This invention combines depthwise separable convolution and dilated convolution to improve classification accuracy and computational efficiency in the case of multiple tree species and small samples, providing a powerful solution for practical applications in the field of tree species classification. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the tree species classification method based on the DSC-DC convolutional neural network of the present invention.

[0032] Figure 2 This is a map showing the distribution of experimental areas and sample plots in this embodiment of the invention;

[0033] Figure 3 This is a schematic diagram of the DSC-DC convolution operation principle in an embodiment of the present invention;

[0034] Figure 4 This is a comparison chart of the prediction effects of this invention with other different classification methods;

[0035] Figure 5 These are detailed diagrams comparing the present invention with other models;

[0036] Figure 6 This is the confusion matrix comparing the present invention with other models;

[0037] Figure 7 This is a structural block diagram of the tree species classification system based on the DSC-DC convolutional neural network of the present invention.

[0038] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] This invention provides a tree species classification method based on a DSC-DC convolutional neural network, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the tree species classification method based on the DSC-DC convolutional neural network of the present invention.

[0041] In this embodiment, the tree species classification method based on DSC-DC convolutional neural network includes the following steps:

[0042] S1: Acquire hyperspectral images of the target area and preprocess the hyperspectral images;

[0043] S2: Establish a DSC-DC classification model consisting of 1 input layer, 7 convolutional layers, 2 pooling layers, 1 flattening layer, 2 fully connected layers, 1 random deactivation layer, and 1 output layer;

[0044] S3: Input the preprocessed hyperspectral image into the established DSC-DC classification model to obtain the output tree species classification results;

[0045] The DSC-DC classification model includes: seven convolutional layers constructed by separable convolutional layers (layers 1, 3, 5, and 8) and depthwise convolutional layers (layers 2, 4, and 7); and two pooling layers constructed by pooling layers (layers 6 and 9).

[0046] Among them, the convolutional layers of the 3rd and 7th layers use dilated convolution with dilation rates of 2 and 4, respectively.

[0047] In a preferred embodiment, step S1 involves acquiring a hyperspectral image of the target area and performing a preprocessing step on the hyperspectral image, specifically including:

[0048] S101: Use a canopy height model to digitize individual tree canopies from hyperspectral images and plot data of the target area so that pixels within the canopy can receive species tags;

[0049] S102: In high-resolution aerial images, create circular labels centered on the center point of each tree and label polygons with the species category, so that each sample label contains only pure pixels of the canopy and does not include the background, soil or other canopy elements.

[0050] In a preferred embodiment, the canopy height model is obtained by processing lidar data of the target area, and the lidar data and the hyperspectral image are configured to be obtained synchronously in the target area.

[0051] In a preferred embodiment, step S1, which involves acquiring hyperspectral images of the target area and performing preprocessing on the hyperspectral images, further includes: performing dimensionality reduction processing on the cubic data of the acquired hyperspectral images using principal component analysis.

[0052] In a preferred embodiment, step S2, establishing the DSC-DC classification model, may include the following specific steps:

[0053] Step 1: Currently, in hyperspectral image classification based on convolutional neural networks (CNNs), most network architectures employ standard convolution operations, which involve multiplying filters element-wise with local regions of the input signal and then summing all the results to generate the output. This leads to a significant increase in computational and parameter complexity as network depth increases, especially when applied to multi-channel hyperspectral data.

[0054] In hyperspectral data classification, the goal of this invention is not only to improve classification accuracy by combining spatial-spectral information for feature extraction, but also to reduce time costs. Therefore, separable convolutions and depthwise convolutions are used alternately to construct convolutional layers, including four separable convolutional layers (layers 1, 3, 5, and 8) and three depthwise convolutional layers (layers 2, 4, and 7) for feature extraction: Layer 1 is a separable convolution, used to perform preliminary feature extraction on the input image; Layer 2 is a depthwise convolution, used to increase the diversity of the model and capture different features; Layer 3 is a separable convolution, performing further feature extraction; Layer 4 is a depthwise convolution, further increasing the diversity of the model and capturing different features; Layer 5 is a separable convolution, further extracting abstract features; Layer 7 is a depthwise convolution, used to enhance the model's sensitivity to larger-scale features; and Layer 8 is a separable convolution, which can be considered the final feature extraction layer, capturing higher-level abstract features in the image. This method effectively extracts multi-scale features while reducing computational complexity, thus maintaining model efficiency while improving classification accuracy. Furthermore, the model's performance and configuration can be further optimized based on specific application requirements.

[0055] Step 2: When performing classification tasks, expanding the model's receptive field helps it better understand the structure and semantics of objects, capture more global features, and improve classification accuracy. In traditional CNN architectures, a direct approach is to use larger convolutional kernels to expand the receptive field. However, using larger convolutional kernels also presents some challenges. Since each convolutional kernel needs to learn a set of weight parameters, as the network depth increases, the number of parameters and computation time also increase, significantly impacting the model's computational efficiency, which poses a challenge to the model's practical applications.

[0056] To address these issues, this invention introduces dilated convolution, which primarily increases the distance between the convolution kernel and the input by adding gaps or holes to the kernel. Compared to traditional convolution, dilated convolution can systematically aggregate multi-scale contextual information without sacrificing resolution, increasing parameters, or computational cost. This improves the accuracy of perception and feature extraction. Furthermore, it effectively expands the receptive field, improving model performance and generalization ability while reducing the risk of overfitting. Therefore, dilated convolution is used in layers 3 and 7 to expand the receptive field, with dilation rates set to 2 and 4, respectively: layer 3, based on separable convolution, uses a convolution operation with a dilation rate of 2 to capture a wider range of contextual information in the feature map, helping to improve the model's receptive field; layer 4, based on depthwise convolution, uses a convolution operation with a dilation rate of 4 to increase the receptive field and extract a larger range of features.

[0057] Step 3: Max pooling layers were applied in layers 6 and 9 to reduce the sampling rate and spatial dimension of the feature maps. In the model design, pooling layers were applied after layers 6 and 9, which extracted rich feature information through convolution operations. This design makes the model focus more on key features rather than over-relying on detailed information. It not only highlights the importance of key features, but also helps to reduce the risk of overfitting, thereby improving the classification performance of the model.

[0058] Step 4: Set the kernel size to 3×3 and set the number of kernels in the next layer to twice that of the previous layer;

[0059] Step 5: Use a flattening layer to flatten the downsampled feature map into a one-dimensional vector;

[0060] Step 6: Perform a nonlinear transformation using a linear rectified function as the activation function for all convolutional and fully connected layers;

[0061] Step 7: Introduce random deactivation with a probability of 0.4 between fully connected layers to reduce overfitting;

[0062] Step 8: Perform multi-class classification using the output layer with the Softmax activation function and output the probability of each class;

[0063] Step 9: Select the optimal model training parameters. Set the batch size to 32, the number of iterations to 100, the window size to 13×13, select the Adam optimizer, and set the optimal learning rate to 0.0001.

[0064] Therefore, this embodiment proposes a tree species classification method based on DSC-DC convolutional neural network. By combining depthwise separable convolution and dilated convolution, it can improve classification accuracy and computational efficiency in the case of multiple tree species and small samples. It provides a powerful solution for practical applications in the field of tree species classification, and realizes scientific, accurate and efficient tree species classification.

[0065] To explain this application more clearly, specific examples of the tree species classification method based on the DSC-DC convolutional neural network are provided below.

[0066] 1) Overview of the study area

[0067] The study area is Teakettle Experimental Forest (TEF), located in the Sierra Nevada Mountains (36°58'00", 119°01'00"), USA. The location of the study area can be found in [link to study area location]. Figure 2 (a) The red dot indicates the location of the study area; (b) Hyperspectral data and label magnification of the red square area of ​​the study area; (c) CHM image and label magnification of the red square area of ​​the study area. The TEF area is approximately 40 hm². 2 The altitude ranges from approximately 1935 to 2630 meters, with an average annual rainfall of about 450 mm. The forest type is mainly mixed coniferous forest, accounting for about 65% of the total forest area, and is mainly distributed at altitudes of 1900 to 2300 meters. The experimental forest has a rich and diverse tree species composition, with the main species being Jeffrey pine (Pinus jeffreyi), white cedar (Abies concolor), redwood (Abies magnifica), sugar pine (Pinus lambertiana), and Korean pine (Pinus contorta).

[0068] 2) Acquisition and processing of remote sensing data

[0069] Airborne lidar and hyperspectral data were collected in July 2017 by the National Ecological Observation Network (NEON) Airborne Observation Platform (AOP) using a DeHavilland DHC-6 Twin Otter aircraft at an altitude of approximately 1000 m. Hyperspectral data was acquired by the AVIRIS next-gen sensor; specific parameters are shown in Table 1. Simultaneously acquired lidar data underwent preprocessing to generate a canopy height model (CHM). The CHM is primarily used to assist in creating labeled data and validating samples in subsequent studies. The original dataset is approximately 16 km long and 1 km wide. Due to hardware limitations, only a portion of the images (approximately 4 km) were used in the experiment.

[0070] Table 1. Hyperspectral data parameters

[0071]

[0072] 3) Acquisition of sample plot data

[0073] The sample plot data was collected by Fricker et al. in September 2017. High-precision Global Positioning System (GPS) was used to obtain individual tree location information, and tree species, diameter at breast height (DBH), and mortality status were recorded in detail. To facilitate species identification from the CHM (Coverage Tree Map), isolated, large-canopy "coverage trees" and smaller trees with clearly separated crowns from surrounding trees were selected during the field survey. The seven tree species and one type of deadwood involved in this embodiment are detailed in Table 2.

[0074] Table 2. Names, codes, and quantities of land categories

[0075]

[0076] 4) Make labels

[0077] First, individual tree canopies were manually digitized from TEF experimental area and plot data using CHM so that pixels within the canopy could receive species labels. Second, circular labels were created in high-resolution aerial images, centered on the center point of each tree, and polygons were labeled with the species category. Each sample label contained only the "pure pixels" of the canopy, excluding the background, soil, or other canopy elements.

[0078] 5) Construction of DSC-DC Convolutional Neural Network

[0079] This embodiment employs a lightweight convolutional neural network. This model uses depthwise separable convolutions to reduce the number of parameters and computational complexity while maintaining good feature extraction capabilities. Dilated convolutions are used to expand the receptive field, enhancing the understanding of image structure. Therefore, the proposed model is named the DSC-DC convolutional neural network, which aims to balance the needs of model performance and computational efficiency.

[0080] Convolutional layers, as the core of CNNs, are key factors determining model performance. The DSC-DC model references Xception and MobileNets in its design. However, unlike directly using depthwise separable convolutions, this invention chooses to alternate between separable and depthwise convolutions. Simultaneously, dilated convolutions are introduced in some convolutional layers to enhance the understanding of image structure. Regarding the determination of kernel size and number, it primarily references most successful CNN network models, setting the kernel size to 33, while adhering to the principle that the number of kernels in each subsequent layer is twice that of the previous layer. Through this design, the DSC-DC model can improve model performance while reducing the number of parameters, thus increasing model efficiency.

[0081] The DSC-DC proposed in this invention for hyperspectral image classification is a pixel-level classification model. A schematic diagram of its convolution operation is shown below. Figure 3 The model consists of an input layer, seven convolutional layers, two pooling layers, one flattening layer, two dense fully connected layers, one dropout layer, and one output layer. The specific workflow is as follows:

[0082] Assume that the hyperspectral image cube data HSI input to the model is M∈S×S×L, where S represents the spatial dimension (width and height) of the input and L represents the number of channels.

[0083] First, this invention applies Principal Component Analysis (PCA) to reduce the dimensionality of M, thereby reducing spectral redundancy and spectral dimension. The result of the dimensionality reduction is an HSI cube of size M ∈ S × S × L. In this model, S is 13 and L is 30. Therefore, the model receives images of a spatial-spectral cube of size M ∈ 13 × 13 × 30 and its corresponding label l as input data through the input layer.

[0084] Next, M∈13×13×30 passes through three separable convolutional layers (layers 1, 3, 5, and 8) and four depthwise convolutional layers (layers 2, 4, and 7) to extract image features. Layers 3 and 7 employ dilated convolutions to expand the receptive field, with spatial dilation rates set to 2 and 4, respectively. Layers 6 and 9 use pooling layers for downsampling to reduce the spatial size of the feature map. Then, fully connected layers flatten the pooling results into a one-dimensional vector. This invention uses the Rectified Linearity Function (ReLU) as the activation function for all convolutional and fully connected layers for nonlinear transformation. To reduce the risk of overfitting, a random deactivation layer with a dropout rate of 0.4 is introduced between fully connected layers. Finally, an output layer with a Softmax activation function is used for multi-class classification, outputting the probability of each class. The model's operational principle is described in [link to model description]. Figure 3 .

[0085] 6) Model parameter setting and optimization

[0086] After determining the network structure, the training parameters of the model were set and tuned. This invention selected a batch size of 32 and set the number of training epochs to 100. Window size is crucial for model training; different window sizes were tried (5×5, 7×7, 9×9, 11×11, 13×13, and 15×15), and the optimal window size was determined to be 13×13. Regarding optimizer selection, SGD and Adam optimizers were compared, and Adam was ultimately chosen as the best option. The choice of learning rate is very important for model training; this invention used a grid search method, selecting an optimal learning rate of 0.0001 from 0.001, 0.003, 0.005, 0.0001, 0.0003, and 0.0005.

[0087] Furthermore, in this embodiment, based on previous experimental experience, dataset size, and the tree classification task, the training, validation, and test datasets are randomly divided. Of the total samples, 80% is used for creating training and validation samples (a 1:1 ratio), and 20% is used for creating test samples. This division provides sufficient training and validation samples to support model training and tuning, while also retaining a relatively large test set for final performance evaluation.

[0088] 7) Comparison of different models

[0089] To evaluate the classification performance and feasibility of the DSC-DC model, it was compared with two classic deep learning models, Inception-V3 and ResNet50, and two machine learning models, KNN and SVM. In the experiments, all models were implemented using Python 3.6. KNN and SVM were based on the scikit-learn framework, while Inception-V3, ResNet50, and DSC-DC were based on the open-source deep learning frameworks TensorFlow and Keras. The operating platform hardware configuration included an NVIDIA RTX A4000 GPU and an Intel(R) Xeon(R) Silver 4216 CPU @ 2.10GHz. The training time based on the deep learning method was the time required to train for 100 epochs.

[0090] To comprehensively evaluate classification accuracy, overall accuracy (OA), average accuracy (AA), and the Kappa coefficient are used as quantitative evaluation metrics for each model. OA represents the ratio of correctly classified pixels to the total number of pixels in all categories. AA represents the average accuracy for each category. The Kappa coefficient measures the consistency between the predicted classification and the ground truth classification, taking into account possible coincidental consistency. In addition, precision, recall, and F1 score are used as evaluation metrics for individual tree species. Higher precision, recall, OA, and F1 score indicate that the predicted values ​​are closer to the true values. The calculation formulas are as follows.

[0091] Overall accuracy:

[0092]

[0093] Average accuracy:

[0094]

[0095] Kappa coefficient:

[0096]

[0097] Accuracy:

[0098]

[0099] Recall rate:

[0100]

[0101] F1 score:

[0102]

[0103] In the formula: N is the total number of samples, n is the number of categories, and x ii Recall is the number of times a class i is correctly predicted. i Let x be the number of recalls for category i. i+ Let t represent the frequency of the true value in category i. TP indicates that both the true and predicted categories are positive, FP indicates that the true category is negative but the predicted category is positive, and FN indicates that the true category is positive but the predicted category is negative.

[0104] 8) Results and Analysis

[0105] (1) Results of tree species classification under optimal parameters

[0106] The samples were trained using the optimal experimental parameters determined above, and the results are shown in Table 3. Among all tree species, redwood, sugar pine, California black oak, and twisted pine showed the best overall performance with no misclassifications. North American cypress and Jeffrey pine all had accuracy, recall, and F1 scores above 0.99, with only two misclassifications. White cedar showed four misclassifications. Dead trees showed only one misclassification. Overall statistics show that accuracy, recall, and F1 scores are all above 0.97, indicating that our model has good classification performance on this dataset.

[0107] Table 3. Tree species classification results of the DSC-DC method under optimal parameters.

[0108]

[0109]

[0110] (2) Comparison results of different classification models

[0111] To further evaluate the classification performance of this model, the DSC-DC model was compared with two classic deep learning models, Inception-V3 and ResNet50, and two machine learning models, KNN and SVM. The quantitative evaluation metrics and time consumption of different classification methods are shown in Table 4, and the classification results are as follows: Figure 4 .

[0112] Compared to machine learning, the DSC-DC model's OA, AA, and Kappa coefficients are significantly higher than those of KNN and SVM. Specifically, compared to SVM, the DSC-DC model improves its OA, AA, and Kappa coefficients by 3.88%, 30.71%, and 0.1566, respectively. Similarly, compared to KNN, the DSC-DC model improves its OA, AA, and Kappa coefficients by 7.55%, 34.09%, and 0.2974, respectively. Analysis of the classification results reveals that KNN and SVM exhibit severe pixel mixing, while the DSC-DC model mitigates this phenomenon to some extent.

[0113] By comparing with deep learning models, it can be seen that the DSC-DC model has certain advantages in classification accuracy. Compared with the Inception-V3 model, the DSC-DC model improves OA, AA, and Kappa coefficients by 0.31%, 0.81%, and 0.0008, respectively. Furthermore, compared with the ResNet50 model, the classification accuracy of the DSC-DC model is not significantly different; its OA and Kappa coefficients achieve the same accuracy of 99.83% and 99.96%, respectively, while its AA is only 0.08% higher than the ResNet50 model.

[0114] In addition to classification accuracy, this embodiment also compares the training time of the models. Under the same epoch, the training time of the ResNet50 model is approximately 5 times that of the DSC-DC model, while the training time of the Inception-V3 model is approximately 3.5 times that of the DSC-DC model. Regarding prediction time, the prediction time of the ResNet50 model is approximately 3 times that of the DSC-DC model, while the prediction time of the Inception-V3 model is approximately 2 times that of the DSC-DC model.

[0115] Furthermore, upon careful observation of the detailed prediction diagram ( Figure 5 Further investigation revealed that the DSC-DC model achieved higher accuracy in extracting tree species boundaries compared to KNN and SVM models. Furthermore, the DSC-DC model demonstrated a significant ability to identify dwarf saplings, a limitation of the Inception-V3 and ResNet50 models.

[0116] Simultaneously, confusion matrices were plotted to further analyze the classification performance of the DSC-DC, ResNet50, and Inception-V3 models. The results are shown below. Figure 6Except for the classification performance of white fir (which was worse than ResNet50) and Jeffrey pine (which was worse than Inception-V3), the DSC-DC model outperformed other networks in classifying tree species for the remaining categories. Redwood, sugar pine, California black oak, twisted pine, and deadwood were all correctly predicted. Due to the small sample size, the ResNet50 model showed serious misclassification of twisted pine, while the DSC-DC model achieved good classification results even with small sample sizes for California black oak and twisted pine. These results demonstrate that the DSC-DC model designed in this invention not only outperforms the other two classic deep learning models in classification accuracy but also has significant advantages in reducing computational burden, greatly saving time costs. It not only improves classification accuracy but is also more efficient in terms of training and prediction time. The tree species classification method proposed in this invention provides a reliable and efficient method for classifying tree species in forests with small sample sizes and multiple tree species.

[0117] Table 4 Comparison of accuracy of different classification methods

[0118]

[0119] Reference Figure 7 , Figure 7 This is a structural block diagram of an embodiment of the tree species classification system based on the DSC-DC convolutional neural network of the present invention.

[0120] like Figure 7 As shown, the tree species classification system based on the DSC-DC convolutional neural network proposed in this embodiment of the invention includes:

[0121] Acquisition module 10, configured to acquire hyperspectral images of the target area and preprocess the hyperspectral images;

[0122] Model building module 20 is configured to build a DSC-DC classification model comprising one input layer, seven convolutional layers, two pooling layers, one flattening layer, two fully connected layers, one random deactivation layer, and one output layer.

[0123] Tree species classification module 30, which is configured to input the preprocessed hyperspectral image into the established DSC-DC classification model to obtain the output tree species classification result;

[0124] The DSC-DC classification model includes: seven convolutional layers constructed by separable convolutional layers (layers 1, 3, 5, and 8) and depthwise convolutional layers (layers 2, 4, and 7); and two pooling layers constructed by pooling layers (layers 6 and 9).

[0125] Among them, the convolutional layers of the 3rd and 7th layers use dilated convolution with dilation rates of 2 and 4, respectively.

[0126] Other embodiments or specific implementations of the tree species classification system based on the DSC-DC convolutional neural network of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0127] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0128] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0129] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0130] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A tree species classification method based on a DSC-DC convolutional neural network, characterized in that, The method includes the following steps: S1: Acquire hyperspectral images of the target area and preprocess the hyperspectral images; specifically including: S101: Digitize individual tree canopies from the acquired hyperspectral images and plot data of the target area using a canopy height model, so that the pixels within the canopy can receive species labels; S102: In the high-resolution aerial images, create circular labels centered on the center point of each tree, and label the polygons with the species category, so that each sample label contains only pure pixels of the canopy and does not contain the background, soil, or other canopies; Principal component analysis was used to reduce the dimensionality of the collected hyperspectral image cube data, and the model input was a 13×13×30 space-spectral cube. S2: Establish a DSC-DC classification model comprising one input layer, seven convolutional layers, two pooling layers, one flattening layer, two fully connected layers, one random deactivation layer, and one output layer; wherein, the kernel size of the seven convolutional layers is 3×3, and the number of kernels in the next layer is twice that of the previous layer; layers 1, 3, 5, and 8 are separable convolutional layers, and layers 2, 4, and 7 are depthwise convolutional layers; layers 3 and 7 perform dilated convolutions with dilation rates of 2 and 4, respectively; layers 6 and 9 apply max pooling layers; the random deactivation layer is configured between two fully connected layers, and the random deactivation layer is configured to randomly deactivate with a probability of 0.4 to reduce overfitting; S3: Input the preprocessed hyperspectral image into the established DSC-DC classification model to obtain the output tree species classification results; The DSC-DC classification model includes: seven convolutional layers constructed by separable convolutional layers (layers 1, 3, 5, and 8) and depthwise convolutional layers (layers 2, 4, and 7); and two pooling layers constructed by pooling layers (layers 6 and 9). Among them, the convolutional layers of the 3rd and 7th layers use dilated convolution with dilation rates of 2 and 4, respectively.

2. The DSC-DC convolutional neural network-based tree species classification method of claim 1, wherein, The canopy height model is obtained by processing lidar data from the target area, and the lidar data and the hyperspectral image are configured to be obtained synchronously in the target area. 3.The DSC-DC convolutional neural network-based tree species classification method of claim 1, wherein, In step S2, the flattening layer is configured to flatten the downsampled feature map into a one-dimensional vector. 4.The DSC-DC convolutional neural network-based tree species classification method of claim 1, wherein, In step S2, a linear rectified function is used as the activation function for all convolutional and fully connected layers to perform a nonlinear transformation.

5. A tree species classification system based on a DSC-DC convolutional neural network, characterized in that, The system for the tree species classification method based on DSC-DC convolutional neural network as described in any one of claims 1-4, the system comprising: The acquisition module is configured to acquire hyperspectral images of the target area and preprocess the hyperspectral images. The model building module is configured to build a DSC-DC classification model that includes one input layer, seven convolutional layers, two pooling layers, one flattening layer, two fully connected layers, one random deactivation layer, and one output layer. A tree species classification module is configured to input preprocessed hyperspectral images into an established DSC-DC classification model to obtain the output tree species classification results. The DSC-DC classification model includes: seven convolutional layers constructed by separable convolutional layers (layers 1, 3, 5, and 8) and depthwise convolutional layers (layers 2, 4, and 7); and two pooling layers constructed by pooling layers (layers 6 and 9). Among them, the convolutional layers of the 3rd and 7th layers use dilated convolution with dilation rates of 2 and 4, respectively.

Images