CN-121999398-A - Unmanned aerial vehicle remote sensing tree species classification method and system integrating cascade canopy attention mechanisms

CN121999398ACN 121999398 ACN121999398 ACN 121999398ACN-121999398-A

Abstract

The invention discloses an unmanned aerial vehicle remote sensing tree species classification method and system integrating a cascade canopy attention mechanism, wherein the method comprises the steps of acquiring unmanned aerial vehicle RGB image data of a campus scene, preprocessing the data to obtain a standardized input image; the method comprises the steps of constructing a classification model fusing a cascade canopy attention mechanism, inputting a standardized input image into the classification model, obtaining an enhanced canopy image through the background suppression module, inputting the enhanced canopy image into the fine granularity feature guiding module, extracting fine granularity canopy features, carrying out channel calibration and feature smoothing on the fine granularity canopy features through the channel alignment neck module to obtain feature mapping with unified dimensions, fusing global canopy semantics and local texture structures in the feature mapping by utilizing the global-local fusion classification head, and outputting tree classification results.

Inventors

XU SHENG
YUAN YINGHUI
YANG YUNFENG
SUN SHUHONG
XIE WENTAO
Wan Jiake

Assignees

南京林业大学

Dates

Publication Date: 20260508
Application Date: 20260305

Claims (8)

1. The unmanned aerial vehicle remote sensing tree species classification method integrating the cascade canopy attention mechanism is characterized by comprising the following steps of: acquiring unmanned aerial vehicle RGB image data of a campus scene, and preprocessing the image data to obtain a standardized input image; Constructing a classification model of a fused cascade canopy attention mechanism, wherein the classification model comprises a background suppression module, a fine granularity feature guiding module, a channel alignment neck module and a global-local fused classification head; Inputting the standardized input image into the classification model, and inhibiting the response of a non-canopy background area through the background inhibition module to obtain an enhanced canopy image; inputting the enhanced canopy image into MobileViT backbone network carrying the fine-granularity feature guiding module, and extracting fine-granularity canopy features; carrying out channel calibration and feature smoothing on the fine-granularity canopy features through the channel alignment neck module to obtain feature mapping with uniform dimension; And fusing the global canopy semantics and the local texture structure in the feature map by using the global-local fusion classification head, and outputting tree classification results.
2. The unmanned aerial vehicle remote sensing tree species classification method integrating the cascade canopy attention mechanism as set forth in claim 1, wherein preprocessing the image data to obtain a standardized input image comprises: cutting and scaling the RGB image of the unmanned aerial vehicle to a fixed size; And carrying out pixel normalization processing on the scaled image, and mapping the pixel value to the [0,1] interval.
3. The unmanned aerial vehicle remote sensing tree species classification method integrating cascade canopy attention mechanisms according to claim 1, wherein the step of suppressing the non-canopy background area response by the background suppression module to obtain the enhanced canopy image comprises the steps of: ; Wherein, the In order to normalize the input image, Representing a depth-separable convolution, A batch normalization is shown and is performed, A 1 x 1 convolution is represented and, Representing the Sigmoid activation function, Representing an element-wise multiplication of the number, To enhance the canopy image.
4. The unmanned aerial vehicle remote sensing tree species classification method integrating cascade canopy attention mechanisms according to claim 1, wherein inputting the enhanced canopy images into a MobileViT backbone network carrying the fine-grained feature guiding module, extracting fine-grained canopy features comprises: Inputting the enhanced canopy image into MobileViT backbone network carrying fine granularity feature guide module, and processing by network layer before preset stage to obtain input feature map of the preset stage; Calculating a spatial attention map by means of a depth convolution from the input feature map; calculating a channel attention map through global average pooling and a full connection layer according to the input feature map; fusing and normalizing the spatial attention map and the channel attention map to obtain a fine granularity mask; Re-weighting the input feature map according to the fine granularity mask to obtain an enhanced feature map; And inputting the enhanced feature map into a tokenization operation, processing and de-tokenizing reconstruction by a transducer encoder, and outputting the fine-granularity canopy features.
5. The unmanned aerial vehicle remote sensing tree species classification method integrating cascade canopy attention mechanisms according to claim 1, wherein the channel alignment neck module performs channel alignment and feature smoothing on the fine-grained canopy features to obtain feature mapping with uniform dimensions, and the method comprises the following steps: ; in the formula, Is a feature of a fine-grained canopy, Number of characteristic channels from Mapping to , A batch normalization is shown and is performed, The representation SiLU activates the function, Feature mapping for unified dimensions.
6. The unmanned aerial vehicle remote sensing tree species classification method based on the cascade canopy attention mechanism, according to claim 1, wherein the global canopy semantics and the local texture structure in the feature map are fused by using the global-local fusion classification head, and outputting tree species classification results comprises: processing the feature map through global average pooling, and extracting global descriptors; Dividing the feature map into K multiplied by K non-overlapping grid areas, carrying out average pooling on each grid area to obtain local descriptors, splicing the local descriptors, and obtaining local representation through linear projection; Splicing the global descriptor and the local representation, and carrying out layer normalization processing to obtain fusion characteristics; and inputting the fusion characteristics into a linear classifier, and processing and outputting tree classification results through a Softmax function.
7. The unmanned aerial vehicle remote sensing tree species classification method integrating cascade canopy attention mechanisms according to claim 1, wherein the training process of the classification model comprises the following steps: Calculating the prediction probability and the loss value of the true tree species label by adopting a cross entropy loss function; And (3) minimizing the loss value through a random gradient descent optimizer, and iteratively updating the model parameters until convergence.
8. An unmanned aerial vehicle remote sensing tree species classification system integrating a cascade canopy attention mechanism is used for implementing the method as set forth in any one of claims 1 to 7, and is characterized by comprising a data acquisition module, a model construction module and a tree species classification module; The data acquisition module is used for acquiring unmanned aerial vehicle RGB image data of a campus scene, preprocessing the image data and obtaining a standardized input image; the model construction module is used for constructing a classification model fusing the cascade canopy attention mechanisms, and the classification model comprises a background suppression module, a fine granularity feature guiding module, a channel alignment neck module and a global-local fusion classification head; the tree classification module is used for inputting the standardized input image into the classification model, restraining the non-canopy background area response through the background restraining module to obtain an enhanced canopy image, inputting the enhanced canopy image into a MobileViT backbone network carrying the fine granularity feature guiding module to extract fine granularity canopy features, carrying out channel calibration and feature smoothing on the fine granularity canopy features through the channel alignment neck module to obtain feature mapping with unified dimension, and fusing global canopy semantics and local texture structures in the feature mapping by utilizing the global-local fusion classification head to output tree classification results.

Description

Unmanned aerial vehicle remote sensing tree species classification method and system integrating cascade canopy attention mechanisms Technical Field The invention belongs to the technical field of unmanned aerial vehicle remote sensing image processing and computer vision, and particularly relates to an unmanned aerial vehicle remote sensing tree classification method and system integrating a cascade canopy attention mechanism. Background Trees in campuses and urban green lands are important components of outdoor space, and play a key role in sun shading, cooling, improving visual comfort, constructing pleasant walking environment and the like. The actual works of refined green land management, daily maintenance, risk investigation, landscape planning and the like all need single-wood-level information support such as tree types, spatial distribution, structural attributes and the like. An Unmanned Aerial Vehicle (UAV) carries an RGB camera, so that single-wood image data of a large-scale campus can be efficiently acquired, and a data base is provided for tree classification. But campus scenes are complex, background interference such as roads, buildings and shadows is serious, the form, color and texture differences of crowns among different tree species are fine, the appearance of crowns of the same tree species also has large variation under different growth environments, and therefore high-precision tree species classification based on unmanned aerial vehicle RGB images faces a great challenge. The existing tree classification methods based on images are mainly divided into three types, namely a method based on manual design features and a traditional classifier, a method based on high-resolution images and a Convolutional Neural Network (CNNs), and a method based on a visual transducer and other large-scale depth models. The method promotes the development of tree classification technology, but has limitations in campus unmanned aerial vehicle scenes, in which most methods do not explicitly design background suppression mechanisms, are difficult to effectively eliminate non-canopy area interference, have insufficient fine granularity distinguishing capability on similar tree species, are difficult to balance similarity and intra-species variability, part of models depend on large backbone networks or complex modules, have large parameter quantity and high calculation cost, are difficult to adapt to deployment requirements of campus scale data sets, and multi-source data fusion can improve classification performance, but has practical difficulties in efficiently integrating multi-source features. The existing method is easy to have the problems of insufficient classification precision, poor robustness and the like under the complex campus background, and is difficult to meet the requirements of precision green land management on tree classification accuracy and integrity. Therefore, there is a need for an unmanned aerial vehicle remote sensing tree classification method capable of effectively suppressing background interference, enhancing canopy characteristics, balancing classification performance and calculating efficiency. Disclosure of Invention In order to solve the technical problems, the invention provides an unmanned aerial vehicle remote sensing tree species classification method and system integrating a cascade canopy attention mechanism, which sequentially realize background suppression, fine-granularity feature guidance and global-local feature fusion by designing a cascade canopy attention module, enhance the signal-to-noise ratio and discriminant of canopy features and improve the accuracy and the robustness of tree species classification in complex campus scenes. In order to achieve the above purpose, the invention provides an unmanned aerial vehicle remote sensing tree species classification method integrating a cascade canopy attention mechanism, comprising the following steps: acquiring unmanned aerial vehicle RGB image data of a campus scene, and preprocessing the image data to obtain a standardized input image; Constructing a classification model of a fused cascade canopy attention mechanism, wherein the classification model comprises a background suppression module, a fine granularity feature guiding module, a channel alignment neck module and a global-local fused classification head; Inputting the standardized input image into the classification model, and inhibiting the response of a non-canopy background area through the background inhibition module to obtain an enhanced canopy image; inputting the enhanced canopy image into MobileViT backbone network carrying the fine-granularity feature guiding module, and extracting fine-granularity canopy features; carrying out channel calibration and feature smoothing on the fine-granularity canopy features through the channel alignment neck module to obtain feature mapping with uniform dimension; And fusing the global canopy semantics and t