Optimized Lightweight Architecture for Drone-Based Solar Panel Contamination Monitoring

This study presents a lightweight and efficient detection model aimed at drone-based solar panel contamination monitoring to address challenges in resource-constrained environments. Our proposed model is built on the MobileViT architecture; it includes a hybrid CNN-Transformer design, localized self-attention mechanisms, and an image preprocessing pipeline. With these optimizations, our model has 6% less parameter count, 8.5% less computational overhead, and 6.1% lower memory usage, while maintaining high detection accuracy. In addition, the preprocessing pipeline utilizes techniques such as Gaussian filtering, histogram equalization, and edge detection. These techniques improve the quality of input images and contribute to more efficient feature extraction. Experimental results strongly suggest that our proposed model outperformed traditional architectures such as DenseNet, ViT, and Swin Transformer. Our proposed model achieved a validation accuracy of 84.1%, which is similar to Swin Transformer and surpassing ViT and DenseNet, but with much lower computational demands. The proposed model is more efficient in solar panel inspections, making it ideal for real-time operability for drones to detect contamination more effectively and cost-efficiently. The proposed model achieved the most optimal balance between accuracy and efficiency; the model demonstrated a 7% improvement in the test accuracy compared to DenseNet with 54% lower FLOPs. This approach offers a sustainable and scalable solution for solar panel maintenance, as it reduces downtime and maximizes energy output. Furthermore, the proposed model’s lightweight and adaptable design is suitable for various drone-based inspection and monitoring applications; future advancements will be feasible particularly in automated and sustainable energy management technologies.