Smart integration of sliding window and vote-based fusion: Advancing UAV-based instance segmentation with YOLOv8 for high-resolution vegetation mapping

Abstract

Vegetation segmentation from high-resolution drone imagery faces fundamental challenges due to scale mismatches between training and inference phases. This issue becomes particularly pronounced when limited high-resolution images are available for model training, necessitating patch-based approaches that create significant scale inconsistencies. The present study addresses these challenges by introducing a spatial consensus voting methodology with confidence-weighted fusion for high-resolution vegetation mapping in the El Ksiba region of Morocco. Unlike previous multi-model ensemble approaches, the proposed method employs a single model generating 10-12 overlapping predictions per pixel, achieving 8.2 × performance recovery from scale mismatch while requiring 3 × less computational cost. Comprehensive evaluation across multiple architectures demonstrates that DeepLabV3+ provides superior precision (0.882) and boundary coherence, while YOLOv8 delivers high recall (0.923) suitable for comprehensive coverage. Contrary to conventional expectations, class imbalance methods consistently degrade deployment performance despite training improvements. This research reveals a focal loss paradox wherein +14.1% training improvement reverses to −1.7% deployment degradation through complex interactions with deployment parameters. Furthermore, minority species benefit substantially more from scale-consistent inference (+55.9% for Lentisque) than from training-time balancing approaches. Experimental results demonstrate effectiveness with mean Intersection over Union of 0.721 and F1-score of 0.826, providing both comprehensive vegetation detection and reliable classification. By resolving scale mismatch problems inherent in training on resized images, this study presents a scalable solution for vegetation segmentation in remote sensing applications, with potential extensions to multimodal sensor integration for enhanced species-specific classification.