LAUSR

Abstract Surface all-wave net radiation (Rn) plays an important role in various land surface processes, such as agricultural, ecological, hydrological, and biogeochemical processes. Recently, remote sensing of Rn at regional and global scales has attracted considerable attention and has achieved significant advances. However, there are many issues in estimating all-sky daily average Rn at high latitudes, such as posing greater uncertainty by surface and atmosphere satellite products at high latitudes, and unavailability of real-time and accurate cloud base height and temperature parameters. In this study, we developed the LRD (length ratio of daytime) classification model using the genetic algorithm-artificial neural network (GA-ANN) to estimate all-sky daily average Rn at high latitudes. With a very high temporal repeating frequency (~6 to 20 times per day) at high latitudes, data from the Moderate Resolution Imaging Spectroradiometer (MODIS) were used to test the proposed method. Rn measurements at 82 sites and top-of-atmosphere (TOA) data of MODIS from 2000 to 2017 were matched for model training and validation. Two models for estimating daily average Rn were developed: model I based on instantaneous daytime MODIS observation and model II based on instantaneous nighttime MODIS observation. Validation results of model I showed an R2 of 0.85, an RMSE of 23.66 W/m2, and a bias of 0.27 W/m2, whereas these values were 0.51, 15.04 W/m2, and −0.08 W/m2 for model II, respectively. Overall, the proposed machine learning algorithm with the LRD classification can accurately estimate the all-sky daily average Rn at high latitudes. Mapping of Rn over the high latitudes at 1 km spatial resolution showed a similar spatial distribution to Rn estimates from the Clouds and the Earth's Radiant Energy System (CERES) product. This method has the potential for operational monitoring of spatio-temporal change of Rn at high latitudes with a long-term coverage of MODIS observations.