MASFNet: Multi-level attention and spatial sampling fusion network for pine wilt disease trees detection
Pine wilt disease (PWD) is a lethal pest and disease that can result in considerable ecological and economic damage. Using deep learning in combination with unmanned aerial vehicle (UAV) orthophoto remote sensing images to detect PWD trees is an advanced and effective method. However, due to the div...
Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-01-01
|
| Series: | Ecological Indicators |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1470160X25000020 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Pine wilt disease (PWD) is a lethal pest and disease that can result in considerable ecological and economic damage. Using deep learning in combination with unmanned aerial vehicle (UAV) orthophoto remote sensing images to detect PWD trees is an advanced and effective method. However, due to the diversity of object information in UAV remote sensing images, most existing algorithms are prone to confusing the background environment and difficult to distinguish highly similar ground objects, resulting in a lot of false detections. To mitigate these challenges, we propose a multi-level attention and spatial sampling fusion network (MASFNet) to reduce false detection problems. Specifically, we designed a Multi-level Attention (MLAttention) module aimed at capturing features through an expanded receptive field to help the model distinguish between the background and the objects of interest. Additionally, we propose a spatial sampling fusion (SSF) module that samples critical sub-regions of features across different layers, generating finer-grained features and thereby enhancing the discrimination between highly similar objects. We utilized UAVs to acquire images and randomly selected images with a resolution greater than 0.6 meters for annotation, creating a dataset of 2360 images for model training and test. Experiments were conducted using seven different models (Faster R-CNN, Cascade R-CNN, CenterNet, FCOS, YOLOX, YOLOv7, and MASFNet), with mean average precision (mAP50) ranging from 0.717 to 0.885. MASFNet demonstrated superior performance in addressing false detection issues in non-forest areas and among highly similar objects. In the improvement experiment of the baseline model, the mAP50 of MASFNet for disease tree detection increased from 0.856 to 0.885. These experimental results indicate that MASFNet can suppress the false detection of PWD and is suitable for disease tree detection tasks in larger forest areas. The code is available at https://github.com/lmctgu/MASFNet. |
|---|---|
| ISSN: | 1470-160X |