CN-122023794-A - Deep learning tree species segmentation method based on few-sample multichannel historical guard sheets

CN122023794ACN 122023794 ACN122023794 ACN 122023794ACN-122023794-A

Abstract

A deep learning tree species segmentation method based on a few-sample multi-channel historical guard sheet is realized based on a deep learning tree species segmentation system of the few-sample multi-channel historical guard sheet, the system is provided with an image acquisition module, a preprocessing module and a tree species segmentation recognition model, the tree species segmentation recognition model is provided with a primary encoder, a secondary encoder, a tertiary encoder, a quaternary encoder and a global feature fusion module, the primary encoder is provided with a convolution input layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, a bridging layer and an upsampling module, the secondary encoder is provided with a dimension raising unit and a first bottleneck block, the tertiary encoder is provided with a second bottleneck block, a first adapter and a first multi-channel attention fusion module, and the quaternary encoder is provided with a third bottleneck block, a second adapter, a fourth bottleneck block and a second multi-channel attention fusion module. The tree species segmentation method has the advantages that tree species segmentation can be effectively and accurately carried out on the historical guard sheets with multiple channels.

Inventors

JIANG CHAOYUAN
TANG CAN
FENG QIANG
CAO XIAOLI
LIU RONGXING

Assignees

重庆英卡电子有限公司

Dates

Publication Date: 20260512
Application Date: 20260114

Claims (7)

1. A deep learning tree species segmentation method based on a few-sample multi-channel history guard slice is characterized by comprising the following steps: Step 1, constructing a deep learning tree species segmentation system based on a few-sample multi-channel historical guard sheet, wherein the deep learning tree species segmentation system is provided with an image acquisition module, a preprocessing module and a tree species segmentation recognition model which are sequentially connected; The tree species segmentation recognition model is provided with a primary encoder, a secondary encoder, a tertiary encoder, a quaternary encoder and a global feature fusion module, wherein the primary encoder is provided with a convolution input layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, a bridging layer and an upsampling module which are sequentially connected; Step 2, the image acquisition module acquires a remote sensing image X to be detected of a woodland area, and then the remote sensing image X is processed by the preprocessing module to obtain a standard remote sensing image C and is transmitted to the tree species segmentation recognition model; Step 3, a primary encoder in the tree species segmentation recognition model acquires the standard remote sensing image C, and a convolution input layer in the primary encoder carries out convolution operation on the standard remote sensing image C to obtain convolution data C0 and transmits the convolution data C0 to a first downsampling layer; step 4, the first downsampling layer performs downsampling operation on the convolution data C0 to obtain first downsampled data C1, and the first downsampled data is transmitted to a second downsampling layer; Step 5, the second downsampling layer performs downsampling operation on the first downsampled data C1 to obtain second downsampled data C2, and the second downsampled data C2 is transmitted to a third downsampling layer and a secondary encoder; Step 6, the third downsampling layer performs downsampling operation on the second downsampled data C2 to obtain third downsampled data C3, and the third downsampled data C3 is transmitted to a fourth downsampling layer and a three-level encoder; Step 7, the fourth downsampling layer performs downsampling operation on the third downsampled data C3 to obtain fourth downsampled data C4, and the fourth downsampled data C4 is transmitted to a bridging layer and a four-level encoder; Step 8, the bridging layer bridges the fourth downsampled data C4, and then upsamples the fourth downsampled data C4 through an upsampling module to obtain upsampled data C9, and transmits the upsampled data C9 to the global feature fusion module; step 9, the second downsampled data C2 is processed by the second level encoder to obtain second level encoded data a, and the second level encoded data a is transmitted to the third level encoder; step 10, the third downsampled data C3 and the second encoded data a are processed by the third level encoder to obtain third level encoded data b, and the third level encoded data b is transmitted to a fourth level encoder; Step 11, the fourth downsampled data C4 and the third coded data b are processed by the fourth-level encoder to obtain fourth coded data d, and the fourth coded data d is transmitted to the global feature fusion module; And step 12, the global feature fusion module performs tree species segmentation identification on the up-sampling data C9 and the four-level coding data d and outputs a tree species segmentation result Y.
2. The method for partitioning deep learning tree species based on few-sample multi-channel history guard slices as set forth in claim 1, wherein said secondary encoder is provided with a dimension raising unit and a first bottleneck block connected in sequence; the three-level encoder is provided with a second bottleneck block, a first adapter and a first multichannel attention fusion module which are sequentially connected; The four-stage encoder is provided with a third bottleneck block, a second adapter, a fourth bottleneck block and a second multichannel attention fusion module which are sequentially connected.
3. The deep learning tree species segmentation method based on the few-sample multi-channel history bathroom piece, which is disclosed in claim 2, is characterized in that the first adapter and the second adapter are consistent in structure, a specific layer perceptron and a shared context perceptron which are sequentially connected are arranged, and a Mish activation function is arranged at the output end of the shared context perceptron; mish the activation function expression is as follows: ; Wherein, the Representing output data of the shared context awareness machine, Representing the hyperbolic tangent function, Representing natural constants.
4. The deep learning tree species segmentation method based on the few-sample multi-channel historical guard sheet is characterized in that the first multi-channel attention fusion module and the second multi-channel attention fusion module are identical in structure, a global average pooling layer, a first full-connection layer, a second full-connection layer and a multi-channel feature fusion layer which are sequentially connected are arranged, the output end of the global average pooling layer is further connected with the input end of the multi-channel feature fusion layer, a ReLU activation function is arranged in the output end of the first full-connection layer, and a Sigmoid activation function is arranged in the output end of the second full-connection layer.
5. The method for deep learning tree species segmentation based on the few-sample multi-channel history bathroom according to claim 4, wherein the multi-channel attention fusion module generates attention weights of each channel through a global average pooling layer, a first full-connection layer and a second full-connection layer, and the expression is as follows: ; the multi-channel feature fusion layer acts the attention weight of each channel on the feature map of the corresponding channel to obtain a fused high-dimensional feature representation, and the formula is as follows: ; Wherein, the , Is a learned weight matrix; attention weight, sign for i channel Representing element-wise multiplication operations, i.e. weights Acting on ith channel profile Is defined by a plurality of pixels; representation of feature graphs A global average pooling operation is performed and, Representing the function of the ReLU activation, Representing the Sigmoid activation function, D represents the total number of channels, and i represents the channel index.
6. The method for deep learning tree species segmentation based on the few-sample multi-channel historical defensive chips, which is characterized in that the tree species segmentation recognition model is obtained by training the following steps: a1, constructing a tree species segmentation recognition model; A2, the image acquisition module acquires a historical guard image set of a woodland area, the preprocessing module cuts each image in the historical guard image set according to a set size, the cut sub-images are arranged into a few sample data set, and then the few sample data set is divided into a training set, a verification set and a test set; And step A3, training the tree species segmentation recognition model by using a training set, continuously optimizing model parameters, verifying the tree species segmentation recognition model by using a verification set, and testing the tree species segmentation recognition model by using a test set to finally obtain the trained tree species segmentation recognition model.
7. The method for segmenting the deep learning tree species based on the few-sample multichannel historical toilet sheets, which is disclosed in claim 6, is characterized in that the historical toilet sheet image set comprises historical toilet sheet images under different seasons and illumination conditions, and the image features of each historical toilet sheet image comprise visible light features, near infrared features and short wave infrared band features.

Description

Deep learning tree species segmentation method based on few-sample multichannel historical guard sheets Technical Field The invention relates to the technical field of tree species segmentation, in particular to a deep learning tree species segmentation method based on a few-sample multi-channel historical guard sheet. Background Tree species segmentation refers to the process of automatically identifying and distinguishing different species of trees from a remote sensing image. In forestry resource management and ecological environment protection, accurate tree species segmentation has important significance in aspects such as evaluating forest health conditions, species diversity protection, carbon sink calculation and the like. However, tree species segmentation for historical defenses presents challenges, particularly tree species diversity and environmental factor variations, which make tree species identification difficult. In China, along with the acceleration of the urban process and the improvement of the ecological environment protection consciousness, the requirement of accurately dividing tree species is urgent. The existing tree species segmentation methods are mainly divided into two types, namely a manual visual interpretation method and an automatic segmentation method based on a computer. The artificial visual interpretation method is a traditional method widely applied to tree species segmentation, and the types and the distribution of tree species are judged by manually observing and interpreting tree species characteristics in a remote sensing image. This approach, while intuitive and suitable for a variety of scenarios, has significant limitations. First, it is highly dependent on the experience and subjective judgment of the operator, and different interpreters may draw different conclusions, resulting in difficulty in ensuring consistency and accuracy of the segmentation results. Secondly, for large-scale remote sensing images, the method is time-consuming and labor-consuming and has low efficiency. Computer-based automatic segmentation methods include rule-based methods, statistical learning methods, and deep learning methods. Compared with manual visual interpretation, the method has the advantages that the automation degree is remarkably improved, and the labor cost can be reduced to a certain extent. However, the conventional rule-based method requires an expert to define a complex rule set in advance, which not only increases the workload in the early stage, but also has limited generalization capability of the rule, and is difficult to cope with complex and changeable tree species characteristics. Statistical learning methods and early machine learning methods, while partially overcoming these problems, often suffer from limited segmentation performance because they are typically based on manually designed features that often do not adequately capture subtle differences in tree species. Thus, the main drawbacks of the prior art include a low degree of automation of the tree species segmentation, especially in the face of multi-channel historical defenses, and a less than ideal efficiency and accuracy of segmentation, especially when dealing with complex and diverse tree species characteristics. Disclosure of Invention The deep learning tree species segmentation method based on the few-sample multi-channel historical guard sheets provided by the invention can realize efficient and accurate tree species segmentation on the historical guard sheets with multiple channels. In order to achieve the above purpose, the invention provides a deep learning tree species segmentation method based on a few-sample multi-channel historical guard sheet, which is characterized by comprising the following steps: Step 1, constructing a deep learning tree species segmentation system based on a few-sample multi-channel historical guard sheet, wherein the deep learning tree species segmentation system is provided with an image acquisition module, a preprocessing module and a tree species segmentation recognition model which are sequentially connected; The tree species segmentation recognition model is provided with a primary encoder, a secondary encoder, a tertiary encoder, a quaternary encoder and a global feature fusion module, wherein the primary encoder is provided with a convolution input layer, a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, a bridging layer and an upsampling module which are sequentially connected; And 2, acquiring a remote sensing image X to be detected of a woodland area by the image acquisition module, processing the remote sensing image X by the preprocessing module to obtain a standard remote sensing image C, transmitting the standard remote sensing image C to the tree species segmentation recognition model, and preprocessing the standard remote sensing image C, wherein the preprocessing comprises the operati