Semi-supervised DNN regression on airborne hyperspectral imagery for improved spatial soil properties prediction

A number of algorithms have been developed for soil organic matter (SOM) or soil heavy metal detection in airborne hyperspectral imagery with high spatial and spectral resolutions. However, to achieve improved land management, the problems of the inconsistent features and low accuracy still need to be solved. In this paper, we propose a novel regression model to estimate the concentrations of SOM, arsenic (As), and chromium (Cr) in soil. Firstly, a hyperspectral unmixing technique is utilized to extract the bare soil pixels. We then combine the absorption depth feature after continuum removal, the original absorption feature, the band ratio feature, and the first-order differential feature, to form a set of features for parameter inversion. To solve the over-fitting problem caused by the small number of samples and the weak expression problem, the semi-supervised deep neural network regression (Semi-DNNR) model is introduced. The experimental were conducted using several datasets collected by HyMap, which is an airborne hyperspectral imaging sensor in VNIR-SWIR spectral range in Yitong county, Jilin province, China. The proposed Semi-DNNR model shows a good performance in this study, with the prediction Rp2 values for SOM, As, and Cr being 0.71, 0.82, and 0.63, respectively. After the spatial distribution map of the soil components of the study area was overlaid with the stream network, which was obtained from the digital elevation model (DEM). It was found that snowmelt, the melting of frozen soil, and surface rainfall can transport SOM to low-lying areas. A similar phenomenon was also observed for As, due to SOM adsorption and dissolved organic matter (DOM) complexation. A comparison of the proposed method with both feature selection methods (competitive adaptive reweighted sampling (CARS), genetic algorithm (GA)) and regression methods (partial least squares regression (PLSR), support vector regression (SVR)) shows that the proposed feature selection method is more robust than the CARS and GA methods. The proposed Semi-DNNR model was found to be at least 18.80% higher in prediction accuracy for As than the SVR or PLSR methods, at least 25.71% higher for Cr, and at least 19.73% higher for SOM.