All-weather VIIRS Land Surface Temperature based on Machine Learning Method

Land Surface Temperature (LST) is a crucial parameter in studies of urban heat islands, climate change, evapotranspiration, hydrological cycles, and vegetation monitoring. Traditionally, LST has been measured at ground stations, offering high accuracy and temporal resolution but limited spatial coverage due to sparse sampling.

Advancements in remote sensing technology have introduced two types of LST data: Thermal infrared (TIR) LST and passive microwave (PMW) LST. TIR LST offers high spatial resolution and accuracy but can be unavailable due to cloud cover. In contrast, PMW LST can penetrate clouds, providing data in all weather conditions, but it has lower spatial resolution and accuracy.

This study focuses on integrating TIR and PMW LST data to obtain high-quality, all-weather LST products, which are required to provide effective scientific support for environmental and climate change research. The research utilizes daily L2 MIRS LST and L3 VIIRS LST (0.009 degrees spatial resolution) products, alongside ancillary data such as Normalized Difference Vegetation Index (NDVI), Shuttle Radar Topography Mission (SRTM) slope, and emissivity.

Machine learning (ML) techniques offer significant improvements over traditional physical models, and ML models benefit from extensive training data. Neural Networks (NN), a subset of ML, are particularly advantageous due to their ability to model complex, non-linear relationships between input and output variables. Additionally, NNs can handle a variety of input features, making them highly effective for tasks involving diverse data sources. Therefore, this study employs the NN method. Specifically, we use MIRS LST, NDVI, SRTM slope, emissivity, and geolocation information as independent variables, with VIIRS LST as the dependent variable. This combination allows us to establish a NN model that generates LST data with high spatial resolution and completeness. The process begins with training the NN model using the comprehensive dataset. The NN learns the intricate relationships between these inputs and the VIIRS LST, improving its predictive accuracy. Once the model is trained, it is evaluated using test data to ensure its reliability and effectiveness.

Evaluation results indicate that the NN model achieves RMSE values of about 4.4 K for daytime and 3.8 K for nighttime all-weather LST, respectively. Overall, the proposed NN method demonstrates promising results, providing all weather LST with improved coverage compared to TIR and higher spatial resolution and improved accuracy than PMW LST data. This enhanced LST data will facilitate the scientific community's research on climate variation, global warming, and other environmental studies.