Comparative Analysis of Convolutional Neural Networks Performance in Remote Sensing for Solar Panel Detection

As the importance of solar energy continues to grow as a pivotal renewable resource, the accurate identification of solar panels within satellite imagery becomes indispensable for both resource assessment and strategic planning. The increasing utilization of solar power signifies a shift towards cleaner and renewable energy sources and serves as a critical measure in mitigating the adverse impacts of climate change, which is predominantly fueled by the widespread use of fossil fuels. This study employs a high-quality, manually labeled dataset comprising 250 satellite images of solar panels in California. Four convolutional neural network (CNN) architectures—DeepLabV3+ VGG16, ResU-Net, UNet3+, and SegNet—are evaluated for their efficacy in detecting solar panels. The models were trained for 50 epochs using a learning rate of 0.0003, the Binary Cross-Entropy loss function, and the Adam optimizer. Performance was assessed using the Dice Coefficient, Recall, and Precision metrics, with DeepLabV3+VGG16 achieving the highest performance with a Dice Coefficient of 94.62% and Recall of 94.23%. These findings provide valuable insights for researchers in selecting appropriate network architectures for accurate semantic segmentation tasks and effective solar panel detection and estimation.