Multi-Sensor Deep Learning Framework for LST Downscaling and Scaling Effect Assessment in Heterogeneous Urban Landscapes

Abstract. Land Surface Temperature (LST) is a critical parameter for urban climate analysis, photovoltaic (PV) planning, and environmental monitoring, yet its effective use is hindered by the coarse spatial resolution of thermal sensors like MODIS. This study introduces a two-stage hierarchical Convolutional Neural Network (CNN) framework that integrates multi-sensor satellite data (MODIS, ASTER, Sentinel-2) to generate fine-scale LST products and quantitatively assess scaling effects across heterogeneous landscapes. In the first stage, MODIS LST (1 km) was downscaled to 360 m, 270 m, and 90 m using Sentinel-2–derived scaling factors and validated against ASTER LST. In the second stage, ASTER LST (90 m) was downscaled to 60 m and 30 m and validated with Landsat-based LST. The framework employed spectral indices, topographic parameters, texture features, and sub-pixel land-cover fractions as scaling factors, capturing both spatial and spectral heterogeneity. Comparative evaluation against XGBoost and DisTrad revealed that CNN consistently achieved the highest determination coefficients (R2 = 0.69–0.87) and the lowest RMSE (1.94–2.34 K) and MAE (1.49–1.8 K) values, confirming its superior capacity to model complex nonlinear thermal relationships. Scaling-effect analysis demonstrated that while accuracy naturally decreases with finer resolutions, the CNN model exhibits strong scale stability and resilience to error propagation, outperforming traditional regression and machine-learning approaches. This hierarchical deep-learning design establishes a new paradigm for multi-sensor LST reconstruction, enabling accurate, scalable, and spatially coherent thermal mapping across diverse terrains. The proposed framework offers a generalizable solution for high-resolution thermal monitoring, PV site optimization, and climate-adaptive urban planning.