The escalating need for efficient biodiversity monitoring motivates this investigation into lightweight deep learning solutions for automated fish species classification. While conventional convolutional neural networks achieve notable accuracy, their computational complexity hinders deployment in resource-limited ecological monitoring scenarios. To address this challenge, this paper presents a neural architecture search-driven framework utilizing differentiable architecture search (DARTS) to automatically design compact models optimized for edge devices. The proposed methodology systematically explores optimal operator combinations with DARTS, such as the size of the convolution kernel and the kinds of pooling operators, etc. Furthermore, the proposed framework incorporates data augmentation strategies to enhance generalization across degraded field images. In the experiment, our derived Model-D achieves 44.62% Top-1 accuracy with 3.39 MB parameters size, representing a 50 times size reduction compared to the large version of ConvNeXt while maintaining comparable accuracy. These findings validate DARTS as a viable strategy for developing fish classification tools that balance computational constraints and performance.