جامعة سومر

Artificial intelligence (AI) has become an essential technology for precision agriculture; however, deep learning models often suffer from architecture-specific bias, in which predictions rely on irrelevant background features rather than meaningful object characteristics. This limitation is particularly evident in UAV-acquired orchard images, which exhibit significant variations in illumination, viewing angle, altitude, and environmental conditions. Such variability may reduce model robustness, explainability, and generalization. This project presents PalmID AI, a novel UAV-assisted Internet of Things (IoT) application for automatically identifying date palm cultivars with reduced AI bias. The proposed framework integrates feature-level fusion between Xception and InceptionResNetV2, enhanced by the Convolutional Block Attention Module (CBAM). Rather than relying on a single convolutional neural network, complementary deep representations extracted from both architectures are fused to suppress architecture-specific bias while improving discriminative feature learning. High-resolution UAV images are processed using the proposed framework, and the resulting fused feature vectors are evaluated with multiple machine learning classifiers. Furthermore, Grad-CAM++ visualization is employed to verify that the proposed model consistently focuses on biologically relevant fruit clusters while minimizing attention to irrelevant background regions. The proposed approach is expected to improve classification accuracy, robustness, explainability, and fairness while reducing prediction variability across different imaging conditions. Consequently, PalmID AI provides a trustworthy AI framework suitable for real-time precision agriculture and intelligent orchard management.