Segmentação de Formigueiros em plantações de cana de açúcar usando imagens de drones

Abstract

This paper presents a semantic segmentation approach for the automatic detection of anthills in aerial orthophotos of agricultural fields. The task is challenging due to the severe class imbalance between background and anthill pixels, the limited number of positive samples, and the visual similarity between anthills and exposed soil. The proposed method employs a U-Net architecture for binary semantic segmentation, combined with mask preprocessing, image normalization, synchronized image-mask transformations, data augmentation for the minority class, and post-processing based on confidence thresholding and connected-region filtering. Different training configurations were evaluated using both detection and segmentation metrics, including Precision, Recall, F1-score, Dice, and Intersection over Union (IoU). The results show that data preparation and class imbalance handling, together with architectural stabilization through normalization and learned upsampling, were key factors in improving model performance. The best configuration prioritized anthill detection, achieving a Recall of 91.7%, an F1-score of 83.1%, and an IoU of 33.8% for the anthill class. Error analysis indicates that false positives are mainly associated with reddish soil and terrain marks, whereas false negatives occur predominantly in small anthills or regions with visual characteristics similar to the background.