Application of MSA-UNet Based on Multi-scale Features and Attention Mechanism Optimization in Drone Photovoltaic Panel Recognition

This study introduces MSA-UNet (Multi-Scale Attention U-Net), an optimized deep learning network for photovoltaic panel recognition in UAV remote sensing images. With rapid advancements in the photovoltaic industry, precise recognition and monitoring of spatial distributions are increasingly essential. Conventional methods face limitations in accuracy due to challenges such as complex backgrounds, variable lighting, and similarities between photovoltaic panels and other ground objects. This paper presents two primary enhancements to the classic U-Net architecture to construct MSA-UNet: (1) a multi-scale feature extraction module that integrates pyramid pooling and dilated convolution, expanding the receptive field to better detect photovoltaic panels of varied scales and improve both large and small object recognition; and (2) a channel attention mechanism, which dynamically adjusts weights during decoding to enhance focus on target areas, thereby effectively reducing background interference and improving segmentation accuracy. Results show that MSA-UNet significantly improves recognition performance, increasing IoU from the base model's 71.5% to 82.0% and Dice coefficient from 73.2% to 84.1%. These findings provide a robust and efficient tool for automated photovoltaic panel detection, supporting optimized layout and resource management in the photovoltaic industry.