UAV-Based Modeling of Land Surface Temperature Using Machine Learning Methods

Land Surface Temperature (LST) is an important climate variable that helps us understand surface heat processes and environmental change. This study focuses on identifying scales at which LST can be reliably modeled using high-resolution RGB and near-infrared (NIR) data as the main input predictors. The approach is based on the well-known negative correlation between the Normalized Difference Vegetation Index (NDVI) and  LST, while vegetation indices represent only one component of the surface energy balance. The study frames LST modeling as a data-driven emulation problem, where surface properties derived from RGB–NIR imagery are combined with concurrent atmospheric and environmental conditions. Several machine learning methods are tested, including Random Forest, XGBoost, LightGBM, and Convolutional Neural Networks, to build an LST emulation framework that links spectral surface information with observed thermal patterns under varying environmental conditions.

The study area is located in the Šumava Mountains in the Czech Republic, a mountain peatland with high ecological value and sensitivity to climate change. Data was collected using a UAV platform between 2025 and 2026, equipped with two sensors: an RGB–NIR camera for surface characterization and a thermal camera used as reference data for surface temperature. These paired multispectral and thermal UAV data form the training basis for the machine-learning models. To ensure the reliability of the models, UAV-derived LST was validated using multiple independent data sources, including in-situ Thermal Infrared (TIR) measurements, near-ground air temperature and humidity monitoring, or air temperature measurements from nearby weather stations.

In addition to spectral variables, the models include several environmental factors that influence surface temperature, such as solar angle, air humidity, soil moisture, wind speed, and canopy height, which act as physical controls on the modeled LST.  A key goal of the study is to test the potential of transfer learning by training the models on data from the Šumava Mountains and evaluating their performance when applied to data from a different season, thereby assessing the temporal robustness of the emulation approach under changing atmospheric and surface conditions.