A Deep Learning-Based Framework for Efficient Path Detection and Planning from Agricultural Drone Imagery

Path detection from drone imagery is a fundamental task in applications such as autonomous navigation, disaster management, environmental monitoring, and precision agriculture. However, the problem remains highly challenging due to variations in altitude, diverse terrain types, occlusions, and fluctuating illumination caused by vegetation and structural obstacles. Conventional approaches often fail to handle these complexities, limiting their adaptability and contextual understanding. To address these issues, this work introduces AeropathNet, a deep learning–based framework designed for efficient path detection and planning in agricultural environments, with direct applicability to IoT-driven swarm robotic systems. The proposed pipeline incorporates preprocessing through Global Contrast Normalization (GCN) and Histogram Equalization (HE), followed by segmentation using a fine-tuned YOLOv5 model. Feature extraction is performed using a deep U-Net architecture, and the most informative features are selected through the Meta-Heuristic Horse Herd Optimization (MH-HHO) algorithm. Final detection and navigation decisions are achieved using a deep ensemble of AlexNet and Inception V3 classifiers. Experimental evaluations on agricultural drone imagery demonstrate that AeropathNet outperforms existing methods, achieving an accuracy of 98.76%, with consistently higher values for precision, recall, F-measure, sensitivity, specificity, and Kappa, along with a significantly reduced loss rate of 0.06%. These results confirm the robustness and efficiency of AeropathNet as a reliable solution for intelligent agricultural navigation.