Revealing scale-dependent controls on possible sunshine duration in mountains: A machine learning simulation of topographic nonlinearities

Possible sunshine duration (PSD) is a key indicator of surface solar radiation and exhibits significant spatial heterogeneity in mountainous regions due to topographic shading. Distributed models based on digital elevation models (DEMs) are widely used to map PSD; however, DEM spatial resolution can introduce substantial discrepancies in simulated results. Traditional statistical methods for analyzing scale effects often fail to capture the nonlinear relationships between topography and PSD. Machine learning offers a promising alternative. Taking Chongqing as a case study, this paper simulates PSD for winter (January) and summer (July) using the Solar Analyst tool applied to DEMs at 30 m, 90 m, 330 m, and 810 m resolutions. An XGBoost model, optimized via Bayesian tuning and interpreted using SHAP (SHapley Additive exPlanations), is employed to analyze the influence of DEM scale on PSD. The results show that: (1) Summer PSD is abundant and relatively uniform, whereas winter PSD is limited and highly variable; both follow an east-low, west-high spatial pattern. Coarser DEM resolutions smooth the spatial pattern from striped to blocky or homogenized distributions. (2) Model accuracy depends on DEM scale and season; as resolution coarsens, R² increases from 0.768 to 0.840 in winter and from 0.674 to 0.852 in summer. (3) The dominant topographic factor governing PSD shifts with scale: in winter, from aspect to terrain undulation; in summer, from terrain undulation to profile curvature, with coarser scales reflecting broader topographic features. (4) Interactions between topographic factors also vary with scale: at 30 m, slope and aspect dominate; at coarser scales, winter PSD is increasingly controlled by elevation and terrain undulation, while summer PSD is shaped by interactions between profile and plane curvature, reflecting the smoothing of local topographic variations. (5) PSD exhibits nonlinear relationships with topographic factors: aspect follows an inverted U-shape, while terrain undulation shows a negative correlation. The proposed XGBoost-SHAP framework effectively reveals scale-dependent mechanisms in PSD simulation and provides theoretical support and practical guidance for multi-scale terrain modeling in mountainous areas, as well as for assessments of solar energy and agricultural resources.