Automated Truffle Crack Detection in Soil Imagery: A Comparative Deep Learning Approach for Precision Agriculture

The labor-intensive process of locating wild desert truffles relies on visually identifying soil cracks indicative of subsurface growth. This study evaluates deep learning’s potential to automate detection by comparing three convolutional neural network (CNN) architectures: a custom model, VGG16 (transfer learning), and ResNet50. Trained on 300 soil images (216 training, 54 validation, 30 test) with real-time geometric augmentation (rotations ±25°, shear ±20%, zoom ±30%), models were tested on a stratified subset of 30 unseen images. The custom CNN achieved 79.6% accuracy (F1=0.727), while VGG16’s transfer learning approach significantly outperformed with 90% accuracy (F1=0.903, AUC=0.938), demonstrating robust feature extraction from limited data. In contrast, ResNet50 catastrophically failed (50% accuracy, 0% specificity), highlighting architectural incompatibility with small-scale crack textures. VGG16’s frozen ImageNet-pretrained layers enabled efficient training (6.7 minutes vs. 9.2 minutes for the custom CNN) and stability under aggressive augmentation, crucial for variable field conditions. Misclassifications (10% error rate) primarily occurred in low-contrast soil textures, emphasizing the need for hybrid architectures integrating attention mechanisms. The results establish transfer learning as optimal for agricultural defect detection, outperforming both manual designs and overly complex models like ResNet50. This work provides a framework for deploying CNNs in precision agriculture, demonstrating that model selection must balance architectural compatibility, training efficiency, and augmentation resilience rather than pursuing depth alone. Practical implications include reduced reliance on manual harvesting and enhanced scalability through mobile deployment. Future directions include multi-modal sensor integration and synthetic data generation to address morphological diversity in truffle cracks.