Construction of a spatially gridded heat flux map based on airborne flux Measurements using remote sensing and machine learning methods

Abstract

Acquiring the relative truth values of land surface sensible (H) and latent (LE) heat fluxes at the regional scale is important for monitoring and simulating evapotranspiration at a global scale using satellite data or process-based models. However, it is difficult to directly obtain these values with fine spatial representation at regional scales using conventional ground-based measurement methods. Therefore, airborne measurements of turbulent flux are highly advantageous. By leveraging airborne measurements of H and LE fluxes collected in the Netherlands in August 2008, in this study we evaluated five machine learning methods for the construction of a regional gridded heat flux map, including artificial neural networks, boosted regression trees, random forest regression, deep neural networks, and support vector regression. The models were trained and tested using a dataset compiled from observations of H and LE, the digital elevation model, MODIS land surface temperature, enhanced vegetation index, and albedo data for each flux footprint. The best performing model was used to construct a regional gridded heat flux map over a case region located in the center of the Netherlands. The results shown that the support vector regression performed better than other models, with R2=0.91,RMSE=9.6W/m2 for H, and R2=0.89,RMSE=26.32W/m2 for LE. The constructed heat flux maps achieved a relative prediction error of ≤32.8% (R2>0.71) for H and ≤43.5% (R2>0.53) for LE compared to aircraft measurements. The constructed heat flux maps were then aggregated for each land cover type, providing individual estimates of source strength (52< H< 66 W/m2 and 108< LE< 218 W/m2) and spatial variability (52< σH< 66 W/m2 and 108< σLE< 218 W/m2) with an ensemble precision of < 6% (1σ). Finally, the limitations and future prospects of the current study were summarized and discussed.