Estimation of surface all-wave net radiation from MODIS data using deep residual neural network based on limited samples

Abstract

Deep learning methods have demonstrated significant success in estimating land surface parameters from satellite data. However, these methods often require large sample sizes for optimal performance, which can be difficult to obtain. This study introduces a novel framework that combines transfer learning (TL) and data augmentation (DA) to improve the performance of a deep learning model, the residual neural network (ResNet), in estimating daily all-wave net radiation (R_{n_daily}) from Moderate Resolution Imaging Spectroradiometer (MODIS) top-of-atmosphere (TOA) observations using limited samples. The framework involves two main steps: first, constructing a robust base model using augmented training samples generated through image rotation in the source region; and second, fine-tuning this base model in target regions with limited local samples. The framework was tested in three regions: the Continental United States (US), Mainland China (MC), and the tropical zone (TR), all with limited ground measurement data. The US was selected as the source region due to its relatively better sample conditions. The results showed that: (1) the ResNet model trained in the US using augmented samples achieved a validated R² of 0.95, RMSE of 14.31, and bias of −0.28 Wm⁻², which outperformed the multi-layer perceptron (MLP) and ResNet models trained with original samples by reducing the validated RMSEs of 2.77 Wm⁻² and 0.80 Wm⁻², respectively; (2) the transferred base model also performed the best in MC and TR, with R² values of 0.86 and 0.66, RMSEs of 22.22 and 25.25 Wm⁻², and biases of 0.22 Wm⁻² and −0.21 Wm⁻², respectively, leading to a decrease in validated RMSE by 3.20, 1.87, and 1.14 Wm⁻² for MC and by 2.32, 1.12, and 0.55 Wm⁻² for TR compared to the MLP and ResNet model trained directly and the ResNet model trained using the augmented samples in these regions, respectively; and (3) the more comprehensive the pre-training sample, the better the framework’s performance in the target domain. However, challenges related to cloud cover and input window size need to be carefully addressed when applying the new framework. Overall, the results highlight the effectiveness of the proposed framework and provide a promising approach for applying deep learning methods with limited samples.