Real-Time Identification of Desert Truffle Cracks via Deep Learning-Based Object Detection in Saudi Arabian Deserts

Economically and nutritionally highly prized, desert truffles are difficult to find as they grow underground and show little visible signs. Often inefficient and environmentally harmful, traditional harvesting techniques depend much on human knowledge. This work presents a creative, automated approach to find cracks in the surface indicating the existence of desert truffles under various Saudi Arabian desert conditions. Two machine learning approaches were employed and compared: a YOLOv2-based object detection model trained entirely from scratch and a transfer-learning model built on pre-existing network weights. A specialized dataset was created by capturing high-resolution images of these tiny surface cracks, manually annotated and subsequently divided into training, validation, and testing subsets. Results showed that the full-network YOLOv2 model outperformed the transfer-learning model by a considerable amount, with an accuracy of 82%, precision of 93%, and recall of 82% in recognizing the minor surface cracks linked with desert truffles. By contrast, the transfer-learning method struggled to generalize to this particular purpose. The results show how deep learning—especially custom-trained YOLOv2 models— can increase real-time identification accuracy for small-scale features such as truffle cracks. Hence boosting truffle harvesting efficiency and sustainability. Future studies will include detailed comparisons with other YOLO variants, extensive data augmentation, and larger datasets.