<p>Crop wild relatives (CWRs) play a crucial role in global food security, as they harbor traits important for crop improvement, such as resistance to diseases, pests, and extreme climatic conditions. However, their conservation and precise identification remain challenging due to morphological similarities with domesticated species, which often result in classification errors and limit their use in breeding programs. This study provides a comparative evaluation of convolutional neural network (CNN) architectures for genus-level identification of CWRs using herbarium images. Both baseline and optimized versions of AlexNet and ResNet50 were assessed, along with additional models including VGG16, MobileNet, and InceptionV3, to expand the scope of comparison. Optimized models integrated architectural refinements and advanced classifiers, including Multi-Layer Perceptron (MLP), which significantly enhanced accuracy. Among the tested architectures, ResNet50 achieved the best performance, reaching a test accuracy of 94.71%, while MobileNet demonstrated competitive efficiency suitable for real-world deployment. The results demonstrate the promise of deep learning in driving progress in biodiversity conservation, digital agriculture, and crop improvement. Future directions include extending the approach to species-level identification, enabling mobile-based applications, and exploring advanced architectures such as DenseNet and EfficientNet.</p>