Evaluation of nine machine learning methods for estimating daily land surface radiation budget from MODIS satellite data

Abstract

The all-wave net radiation (Rn) at the land surface represents surface radiation budget and plays an important role in the Earth's energy and water cycles. Many studies have been conducted to estimate from satellite top-of-atmosphere (TOA) data using various methods, particularly the application of machine learning (ML) and deep learning (DL). However, few studies have been conducted to provide a comprehensive evaluation about various ML and DL methods in retrieving. Based on extensive in situ measurements distributed at mid-low latitudes, the corresponding Moderate Resolution Imaging Spectroradiometer (MODIS) TOA observations, and the daily from the fifth generation of European Centre for Medium-Range Weather Forecasts Reanalysis 5 (ERA5) used as a priori knowledge, this study assessed nine models for daily estimation, including six classic ML methods (random forest -RF, adaptive boosting - Adaboost, extreme gradient boosting -XGBoost, multilayer perceptron -MLP, radial basis function neural network -RBF, and support vector machine -SVM) and three DL methods (multilayer perceptron neural network with stacked autoencoders -SAE, deep belief network -DBN and residual neural network -ResNet). The validation results showed that the three DL methods were generally better than the six ML methods except XGBoost, although they all performed poorly in certain conditions such as winter days, rugged terrain, and high elevation. ResNet had the most robust performance across different land cover types, elevations, seasons, and latitude zones, but it has disadvantages in practice because of its highly configurable implementation environment and low computational efficiency. The estimated daily values from all nine models were more accurate than the corresponding Global LAnd Surface Satellite (GLASS) product.