Deep Learning-Based Analysis for Intelligent Identification of Continuously Illuminated Regions at Lunar South Pole

This study examines the lunar south pole environment centered on the connecting ridge between Shackleton and de Gerlache craters, leveraging illumination simulations from November 1st, 2026, to February 28th, 2027, to generate a dynamic illumination dataset with a spatial resolution of 20 m/pixel and a temporal resolution of 1 hour. A novel deep learning framework is introduced to identify regions with at least three consecutive days of continuous illumination. The framework integrates an enhanced deep convolutional neural network to extract illumination features from individual temporal frames with a bidirectional gated recurrent unit to model temporal dependencies across sequences. A dual temporal-spatial attention mechanism evaluates illumination continuity, assigns spatiotemporal weights, and emphasizes critical features, while an upsampling output head produces high-resolution maps. The model demonstrates superior performance in extracting stable features from sequential data, enabling precise detection of three-day persistent illumination regions at a 20 m resolution. When coupled with an optimized A* algorithm, the framework supports efficient rover path planning.