OpenPineDet: An open-set detection framework for the detection of dead pine wood in forests

Early and accurate detection of dead pine wood (DPW) is essential to mitigate the spread of pine wilt disease (PWD), a destructive affliction of pine forests. Deep learning-based object detection methods have been widely employed for DPW detection. However, these approaches typically assume that both the training and test datasets contain only instances from a single, known class (i.e., DPW). This closed-set assumption often results in unknown targets in the test data being misclassified as DPW. To address this limitation, we collected high-resolution drone imagery of pine forests, resulting in a dataset of 7,733 images (the GSD ranged from 10.4 to 15.5 cm/pixel for flight altitudes between 800 and 1200 m) comprising 29,575 samples of DPW and 7,334 samples of unknown categories. We then introduced an open-set object detection framework, OpenPineDet (OPDet), designed to accurately identify DPW while recognizing other similar objects (e.g., bare soil and houses) as unknown categories. OPDet integrates two novel modules: (1) a Bottleneck Channel Attention Network (BNCA) that leverages the Efficient Channel Attention (ECA) mechanism to enhance the model’s focus on DPW targets; and (2) a Dual Contrastive Learning (DCL) module, which combines Instance-Level Contrastive Learning (ICL) with Focal Loss (FL) to improve detection accuracy of hard-to-classify DPW instances. Experimental evaluations demonstrate that OPDet outperforms existing methods, achieving an Average Precision (AP K ) of 92.67 % and a Recall (R K ) of 97.24 %, while reducing Wilderness Impact (WI) and Absolute Open-Set Error (AOSE) to 1.80 and 242, respectively. Additionally, the proposed model demonstrates stable performance across different scenarios (such as simple scenes with only pine trees and complex scenes with other unknown categories), and exhibits better regional generalization capabilities. These results indicate that OPDet effectively reduces the misclassification of unknown targets as DPW, highlighting its effectiveness for open-set detection methodologies in forest management applications.