Combination of Hyperspectral and Machine Learning to Invert Soil Electrical Conductivity

An accurate estimation of soil electrical conductivity (EC) using hyperspectral techniques is great significance for understanding the spatial distribution solutes and salinization. Although spectral transformation has been widely used in data pre-processing, performance different pre-processing (or combination methods) on models same set still ambiguous. Moreover, extremely randomized trees (ERT) light gradient boosting machine (LightGBM) are new learning algorithms with good generalization (soil moisture above-ground biomass), but less studied estimating salinity visible near-infrared spectra. In this study, 130 EC data, measured topographic factors, conventional indices such as Salinity Index 1, two-band (2D) ratio indices, were introduced. The five methods standard normal variate (SNV), detrend (SNV-DT), inverse (1/OR) (OR original spectrum), inverse-log (Log(1/OR) fractional order derivative (FOD) (range 0–2, intervals 0.25) performed. A (GBM) was to select sensitive parameters. Models (extreme (XGBoost), LightGBM, random forest (RF), ERT, classification regression tree (CART), ridge (RR)) inversion model validation. results reveal that two-dimensional correlation coefficient highlighted more effectively than one-dimensional. Under SNV second derivative, increased by 0.286 0.258 compared one-dimension, respectively. 13 characteristic factors slope, NDI, SI-T, RI, profile curvature, DOA, plane SI (conventional), elevation, Int2, aspect, S1 TWI provided 90% cumulative importance GBM. Among six models, ERT performed best simulation (R2 = 0.98) validation 0.96). showed among from reference data. kriging map based a close relationship Our study selected effective (SNV 2 derivative) one- correlation, important inversion. This provides theoretical support quantitative monitoring salinization larger scale remote sensing techniques.