WSPNET:Weakly Supervised Learning With Pesudomask For Tea Pest Detection

This article tackles the pressing issue of pest detection in tea farming, focusing on two significant pests: the tea thrips (Scirto thrips dorsalis) and the looper (Hypo Sidra Talca). These pests pose serious risks to the health and yield of tea crops. While tea's economic significance is well recognized, much of the existing research has primarily cantered on larger pests in more controlled environments, leaving a noticeable gap in reliable detection methods for smaller pests that thrive in the dense and diverse settings of tea plantations. To address this imbalance, we propose utilizing advanced deep learning models, particularly the UNet and Enriched UNet architectures, to improve the accuracy of pest detection and minimize false positives. Our approach involves creating a robust segmentation model that integrates attention mechanisms, residual connections, and tailored loss functions designed to tackle class imbalance and enhance boundary detection. By harnessing high-resolution imaging and sophisticated artificial intelligence techniques, this research aims to support sustainable pest management in tea cultivation, ultimately enhancing both crop protection and economic sustainability.