Predicting soil moisture content over partially vegetation covered surfaces from hyperspectral data with deep learning

• DOI: 10.1002/saj2.20193
• Publication Date: 2020-11-10T10:43:10Z
• Authors: Fangfang Zhang; Shiwen Wu; Jie Liu; Changkun Wang; Zhiying Guo; Aiai Xu; Kai Pan; Xianzhang Pan
• Keywords: 0401 agriculture, forestry, and fisheries; 04 agricultural and veterinary sciences; 15. Life on land

Abstract Previous studies for retrieving soil moisture content (SMC) from visible and near‐infrared hyperspectral data over vegetation‐covered surfaces using spectral unmixing, non‐negative matrix factorization, albedo/vegetation coverage in trapezoid spaces have required mass preprocessing offered only limited improvements prediction accuracy. Recently, deep learning has triggered some properties because of its automatic feature extraction high In this study, a simulation experiment with different vegetation coverages, SMCs, types were acquired. Deep models, one‐dimensional convolutional neural network (1D‐CNN), long short‐term memory (LSTM) are proposed to predict SMC. The results showed that two models achieved excellent predictions (residual deviation [RPD] > 2.5) the unpreprocessed mixed spectra partial least squares regression (PLSR) had good (RPD = 1.88). 1D‐CNN ( R 2 p .91) LSTM .90) significantly outperformed PLSR .72), which demonstrated could improve SMC partially surfaces. However, when bare spectra, accuracy was commensurate, whether through 1D‐CNN, LSTM, or models; additionally, better performance on all than spectra. These indicated no advantage smaller datasets. We also found affected by type but still very good. effective predicting large datasets acquired complex surface conditions.