Impact of the Loss Function on the Performance of U-Net Models for Leaf Miner Detection in Meloeiro

Brazil is among the world's largest exporters of melon, which drives the search for technologies that enhance the quality and productivity of this crop. In this context, this study proposes the analysis of more effective loss functions for a semantic segmentation model capable of detecting and delineating regions affected by infestations, focusing on leaf miner, a common pest in melon cultivation. To this end, a novel dataset was developed with images collected under real field conditions. Pixel-based and region-based loss functions were evaluated using a U-Net architecture. The results indicated superior performance of region-based loss functions, with emphasis on Log-Cosh Dice Loss, which achieved approximately 69% F1-score and 53% IoU. The model presents satisfactory results and generalization potential, and can be extended to other diseases in agricultural crops.