In arid and semi-arid regions, water availability strongly constrains watermelon production, making timely water-stress assessment essential for precision irrigation. This study proposes a thermal-imaging-based approach to classify watermelon water-stress status using deep learning. A low-cost thermal camera was used to acquire leaf thermal images under different soil-moisture conditions, yielding a dataset of 2168 images after augmentation. To move beyond single-model evaluation, we benchmarked representative architectures spanning classical convolutional neural networks, lightweight networks, Vision Transformers (ViTs), and modern convolutional architectures. Model performance was assessed not only by classification accuracy but also by deployment-oriented efficiency metrics, including parameter count, computational complexity (GFLOPs), and inference latency. Among the tested models, EfficientNet-B0 achieved the best overall trade-off: accuracy and F1 score both reached 0.99 while maintaining a compact size (5.3 M parameters), low computation (0.39 GFLOPs), and fast inference (8.81 ms). These results indicate that EfficientNet-B0 is well-suited for practical thermalimage-based diagnosis of watermelon water stress. Overall, this work provides a systematic comparison of deep learning models for crop water-stress classification from thermal imagery and offers guidance for model selection in field-deployable irrigation management systems.