Abstract
Camels are resilient animals that play a crucial role in arid ecosystems and desert communities. However, distinguishing between visually similar camel breeds-particularly among Arabian camels-remains a challenging task. This paper introduces a novel image dataset containing 1,620 images of Arabian and Non-Arabian camels, collected and annotated by the authors. Based on this dataset, we propose an explainable multi-stream deep learning architecture for fine-grained classification using a hierarchical adaptive framework. The model operates in two stages: (1) binary classification to distinguish Arabian camels from Non-Arabian camels, and (2) multi-class classification to identify five distinct Arabian camel breeds-Homor, Majaheem, Sofor, Waddah, and Shaele. The multi-stream design enables the model to process global context, local features, and semantic cues in parallel, significantly improving feature extraction and classification accuracy. Among several evaluated CNNs-including DenseNet121, MobileNetV2, InceptionV3, and ResNet50-DenseNet121 achieved the highest performance, yielding 98% accuracy in binary classification and 76% in multi-class classification. To address class imbalance and enhance generalization, we applied online data augmentation, class-balanced focal, and the Adam optimizer. For interpretability, Grad-CAM was integrated to visualize key decision-making regions, enhancing transparency and trust in AI-based predictions. Despite challenges such as visual similarity among certain breeds and dataset imbalance, our approach demonstrates promising results and establishes a foundation for future work in automated camel breed identification and livestock management.