A method to estimate leaf area index from VIIRS surface reflectance using deep transfer learning

Abstract

The leaf area index (LAI) retrieval methods based on traditional neural networks require a large number of training samples constructed from remote sensing data or simulation data using radiative transfer models. Furthermore, the training samples for the neural networks are sensor-specific. Therefore, the existing training samples for a sensor cannot be directly applied to estimate LAI values from remote sensing data acquired by another sensor. In addition, a large number of currently available LAI ground measurements (considered as “true values”) are not used to construct training datasets of the neural networks to further improve the accuracy of the retrieved LAI values. In this study, a method based on deep transfer learning is proposed to retrieve LAI values from the Visual Infrared Imaging Radiometer Suite (VIIRS) surface reflectance data based on the existing training dataset established from the Moderate Resolution Imaging Spectroradiometer (MODIS) surface reflectance data and the fused LAI values of the MODIS and Carbon cYcle and Change in Land Observational Products from an Ensemble of Satellites (CYCLOPES) LAI products. A transferable model is constructed by using a deep belief network (DBN) composed of a restricted Boltzmann machine (RBM) and a back propagation (BP) network. The DBN is pretrained by the existing training dataset, and the hyperparameters of the DBN are determined by using a Bayesian optimization algorithm. Then, the RBM parameters of the pretrained DBN are frozen, and the BP parameters are fine-tuned by a small sample dataset composed of the LAI ground measurements and the corresponding VIIRS surface reflectance data. Finally, the VIIRS surface reflectance data are inputted into the fine-tuned DBN to estimate LAI values. The retrieved LAI values were cross-compared with the MODIS and VIIRS LAI products and were directly evaluated with LAI ground measurements at the Ground-Based Observations for Validation (GBOV) and Implementing Multi-Scale Agricultural Indicators Exploiting Sentinels (IMAGINES) sites. The results demonstrate that the retrieved LAI values at these sites with different vegetation types show reasonable seasonality and the spatial distributions of the retrieved LAI values are similar to those of the MODIS and VIIRS LAI products. The direct validation results show that the LAI values retrieved by the DBN with fine-tuned BP parameters (RMSE over the USA: 0.48, RMSE over Europe: 0.71) are obviously superior to the LAI values retrieved by the DBN without fine-tuned BP parameters (RMSE over the USA: 0.70, RMSE over Europe: 0.98). This study demonstrates that deep transfer learning can effectively retrieve LAI values from the VIIRS surface reflectance data with limited LAI ground measurement samples and the existing training dataset.