Path Aggregation Network based WATT-EffNet for Unmanned Aerial Image Classification

High-resolution aerial images captured by UAVs are critical in various real-world applications. However, effective classification of these images is challenging due to scale variation, occlusions, and complex scene structures. Existing deep learning models often face a trade-off between computational efficiency and classification accuracy. To address this issue, a novel Path Aggregation Network-based Wider Attention EfficientNet (PANet-WATT-EffNet) is proposed. PANet-WATT-EffNet employs EfficientNet as a lightweight backbone, combined with wider attention layers to capture salient regions. PANet is used for multi-scale feature fusion. PANet-WATT-EffNet’s design improves the extraction of fine-grained and global features, enabling accurate recognition of small and complex objects in aerial imagery. Experimental evaluation on UAV benchmark datasets shows that the model achieves 97.71% accuracy. A significant gain is observed in F-measure, MCC, and reduced RMSE values of PANet-WATT-EffNet, while also lowering computational time compared to existing methods. The results confirm the robustness of the approach in handling diverse aerial imagery. The lightweight architecture further supports deployment on resource-constrained edge devices, making it suitable for applications in precision agriculture, urban infrastructure monitoring, disaster management, and defence surveillance.