Inversion of Different Cultivated Soil Types’ Salinity Using Hyperspectral Data and Machine Learning

Abstract

Soil salinization is one of the main causes of global desertification and soil degradation. Although previous studies have investigated the hyperspectral inversion of soil salinity using machine learning, only a few have been based on soil types. Moreover, agricultural fields can be improved based on the accurate estimation of the soil salinity, according to the soil type. We collected field data relating to six salinized soils, Haplic Solonchaks (HSK), Stagnic Solonchaks (SSK), Calcic Sonlonchaks (CSK), Fluvic Solonchaks (FSK), Haplic Sonlontzs (HSN), and Takyr Solonetzs (TSN), in the Hetao Plain of the upper reaches of the Yellow River, and measured the in situ hyperspectral, pH, and electrical conductivity (EC) values of a total of 231 soil samples. The two-dimensional spectral index, topographic factors, climate factors, and soil texture were considered. Several models were used for the inversion of the saline soil types: partial least squares regression (PLSR), random forest (RF), extremely randomized trees (ERT), and ridge regression (RR). The spectral curves of the six salinized soil types were similar, but their reflectance sizes were different. The degree of salinization did not change according to the spectral reflectance of the soil types, and the related properties were inconsistent. The Pearson’s correlation coefficient (PCC) between the two-dimensional spectral index and the EC was much greater than that between the reflectance and EC in the original band. In the two-dimensional index, the PCC of the HSK-NDI was the largest (0.97), whereas in the original band, the PCC of the SSK400 nm was the largest (0.70). The two-dimensional spectral index (NDI, RI, and DI) and the characteristic bands were the most selected variables in the six salinized soil types, based on the variable projection importance analysis (VIP). The best inversion model for the HSK and FSK was the RF, whereas the best inversion model for the CSK, SSK, HSN, and TSN was the ERT, and the CSK-ERT had the best performance (R2 = 0.99, RMSE = 0.18, and RPIQ = 6.38). This study provides a reference for distinguishing various salinization types using hyperspectral reflectance and provides a foundation for the accurate monitoring of salinized soil via multispectral remote sensing.

UAV hyperspectral analysis of secondary salinization in arid oasis cotton fields: effects of FOD feature selection and SOA-RF

“…There are many studies that use machine learning methods for soil salinity prediction ( Zhu et al., 2022 ), utilized RF model R 2 to reach 0.815 in a study of soil salinity inversion of Abbey Lake Oasis based on UAV hyperspectral imagery; ( Jia et al., 2022 ) used an RF model with an R 2 of 0.93 in a study of salinity inversion for different cultivated soil types based on hyperspectral data and machine learning; ( Hu et al., 2019 ) reached an R 2 of 0.94 in quantitative soil salinity estimation using RF model predictions based on UAV hyperspectral data; ( Cui et al., 2023 ) used unmanned aerial multispectral remote sensing and three machine learning algorithms to invert the soil salinity of farmland at different depths under crop cover, and the optimal prediction model had an R 2 of 0.775. After comparison, the accuracy of the SOA-RF model used in this study exceeds that of the previous models used ( Zhao et al., 2022 ).…”

Low-Cost Optical Sensors for Soil Composition Monitoring

“…As in the previous case, spectroscopic methods are also used to determine soil salt content. In [ 44 , 45 ], the conjunction of NIR and Vis spectra were used as inputs for evaluating the electroconductivity (EC) of soil and soil saline content (SCC), and the obtained R 2 reached 0.99 [ 44 ] and 0.87 [ 44 , 45 ] for the best combination. In some cases, image processing is also combined with spectroscopic information [ 35 ].…”

“…Considering the different studied samples, in all cases, the authors used natural samples [ 30 , 35 , 43 , 44 , 45 , 46 ]. Only in [ 47 ] were the analyzed samples artificially generated, as in our case.…”

“…The number of classes for the classification algorithms ranged from three [ 47 ] to five [ 35 , 46 ]. There is a wide variety of classification and regression methods such as CNN [ 35 ], ANN [ 46 ], RF [ 35 , 43 , 44 , 46 ], support vector machine (SVM) [ 35 , 46 ], Gated Recurrent Units (GRU) [ 35 ], and Native Bayesian (NB). None of the included ones used DA.…”

“…None of the included ones used DA. Among the most used regression tools, PLSR [9,43,44] is the most used, while others such as random forest regression (RFR), support vector regression (SVR), Gradient-Boosted Regression Tree (GBRT), Multilayer Perceptron Regression (MLPR), and Least Angle Regression (Lars) are used only in [45]. Among surveyed examples of sensing elements for soil sensor characteristics, the proposed ones offered a series of advantages.…”