Fuzzy EfficientDet: An Approach for Precise Detection of Larch Infestation Severity in UAV Imagery Under Dynamic Environmental Conditions

Abstract

In this paper, a novel deep learning framework, fuzzy EfficientDet, is proposed to address the challenge of accurately detecting larch infested by Coleophora laricella pests in UAV imagery, where the key innovation is the incorporation of the fuzzy spatial attention mechanism (FSAM), which can effectively deal with the problem of model uncertainty due to the complexity of environmental transformations and image features. First, this study designs and implements a Global-Local Squeeze-and-Excitation Module, which profoundly integrates global and local feature information, realizes the dynamic adaptation of the importance of feature channels in the EfficientNet, and thus improves the overall feature expression efficiency of the network. Second, this study constructed a dense Bi-FPN architecture, which adds a dense connection structure to the original Bi-FPN to enhance the modeling accuracy for small targets and long-range spatial dependencies. Finally, this study develops the fuzzy spatial attention mechanism (FSAM), which can effectively mitigate the unstable performance of the EfficientDet in the face of image feature fluctuations triggered by changes in lighting conditions and seasonal effects. Experiments demonstrate that the proposed fuzzy EfficientDet model shows superior performance compared to the traditional SSD, Faster R-CNN, YOLO V5, and the unimproved EfficientDet target detection method on the Swedish Forest Agency (2021) dataset, with its mAP as high as 94.29%. This result demonstrates that fuzzy EfficientDet provides an efficient and reliable solution when dealing with the task of target detection in UAV images, especially in dealing with environmental uncertainty and complex feature extraction.