Oil Palm Detection from High Resolution Remote Sensing Imagery with Improved YOLO Deep Learning Frameworks

Abstract

Purpose: Most existing oil palm detection studies rely on standard You Only Look Once (YOLO) models without architectural modifications, limiting their detection performance. This study aims to improve the YOLO11 model for oil palm detection by integrating various attention mechanisms.Research Method: A total of 53,113 annotations of WorldView-2 and UAV datasets were used for training (80%), validation (10%), and testing (10%). An external validation dataset was included to evaluate model generalizability to different tree crown sizes and spatial characteristics (young oil palm, mature oil palm, mixed land cover). To improve the YOLO11 model, attention modules such as Coordinate Attention (CA), Efficient Channel Attention (ECA), and Convolutional Block Attention Module (CBAM) were integrated into the neck component. Model evaluation was conducted in terms of accuracy assessment and processing times.Findings and Values: For the primary dataset, YOLO11-CBAM achieved the highest mAP50-95 for validation (0.605) and testing (0.611). For the external validation dataset, YOLO11-CBAM also achieved the highest Recall and F1-score for both the young oil palm region (Recall: 0.988, F1-score: 0.992) and the mixed land cover region (Recall: 0.924, F1-score: 0.953). However, YOLO11-ECA achieved the highest accuracy across all metrics for the mature oil palm region. This study demonstrates that adding attention modules to the YOLO11 architecture improves detection accuracy and generalization despite the slightly increased inference time. The attention-enhanced models outperformed the baseline model and benchmark detectors (Faster R-CNN, RetinaNet) in oil palm detection across different environmental conditions.