A practical machine learning approach to retrieve land surface emissivity from space using visible and near-infrared to short-wave infrared data

• Crucial variables for land surface emissivity retrieval are analyzed. • The machine learning model exhibits exceptional performance in accurately retrieving land surface emissivity using visible and near-infrared as well as short-wave infrared data. • Models driven by pixel-scale products demonstrate robustness and universality across diverse temporal and spatial domains. • The transferability of the method on other satellites has been validated, affirming its versatility and applicability. Land surface emissivity (LSE) is a crucial variable in thermal infrared (TIR) remote sensing, providing unique information about the land surface across different channels. It is essential for applications such as surface energy budget estimation, resource exploration, and land cover change monitoring. However, current methods for retrieving LSE have certain limitations in terms of applicability or accuracy levels. Furthermore, the relative importance of various parameters in LSE retrieval studies remains unclear. To address these challenges, a practical and transferrable method has been proposed to retrieve LSE of different TIR channels using machine-learning technique. The proposed method uses visible and near-infrared (VNIR) as well as short-wave infrared (SWIR) data at the pixel scale to analyze key parameters for LSE retrieval and to estimate LSE for channels centered around 8.6 μm, 11.0 μm and 12.0 μm. Importance analysis identified crucial variables for LSE retrieval, including reflectivity in channels of SWIR3 (∼2.13 μm), RED (∼0.66 μm) and BLUE (∼0.47 μm), as well as the Enhanced Vegetation Index (EVI), Normalized Difference Water Index (NDWI), Normalized Difference Built-up Index (NDBI) and the view zenith angle (VZ). Compared to the data used in existing methods, the core variables offer a more comprehensive representation of surface information, potentially enhancing both the accuracy and usability of the proposed method. Using these core variables, LSE was retrieved across eleven study areas through a machine learning method. Cross-validation with MODIS products showed that the Root Mean Square Error (RMSE) of the estimated LSE is 0.02 for the channel around 8.6 μm, and 0.01 for the channels around 11.0 μm and 12.0 μm, respectively. Direct-validation with in-situ measurements also demonstrated impressive retrieval accuracies in sandy areas. Furthermore, the model trained using 2019 data exhibited high retrieval accuracy when applied to data from 2017, highlighting its transferability across different time periods. Additionally, the proposed method produced promising results for LSE estimation using Landsat 8 imageries, indicating its potential for generating emissivity products from satellites with high spatial resolution but limited TIR channels.