CNN-Based Downscaling of Land Surface Temperature and Scaling Effect Analysis Using Multi-Sensor Satellite Data: A Case Study in L’Aquila, Italy

Abstract. High-resolution Land Surface Temperature (LST) is vital for applications such as photovoltaic (PV) plant planning, urban heat assessment, and environmental monitoring. However, the coarse spatial resolution of sensors such as MODIS limits their suitability for local-scale analyses. To address this limitation, this study presents a deep learning Convolutional Neural Network (CNN)-based framework for downscaling Land Surface Temperature (DLST) from MODIS data at coarse resolution to finer resolutions of 270 m, 90 m, 60 m, and 30 m over L’Aquila, Italy, using scaling factors exclusively derived from Sentinel-2 optical data. Validation is performed against high-resolution reference datasets: ASTER 90 m and its upscaled LST (ULST) aggregated to 270 m for evaluating DLST at 90 m and 270 m, and Landsat 30 m and its ULST aggregated to 60 m for evaluating DLST at 30 m and 60 m. A comprehensive quantitative assessment is carried out using regression analysis, coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and p-value testing to rigorously evaluate accuracy and assess scaling effects. The results were compared against ASTER- and Landsat-derived LST, showing that the CNN-based DLST achieved RMSE values ranging from 2.11 K to 2.94 K and MAE values from 1.62 K to 2.44 K across resolutions, maintaining robust agreement with both high-resolution references and their aggregated ULST counterparts. A key finding from the scaling effect assessment is that while DLST remains statistically consistent with reference LST down to 60 m and 90 m, performance at 30 m shows a clear decline in agreement (p-value = 0.31), indicating that scaling effects become more influential at very fine resolutions and require careful consideration when interpreting high-resolution outputs. The findings demonstrate that CNN-based DLST can generate fine-scale LST products suitable for PV site selection, operational monitoring, and broader geospatial applications.