A novel method for detecting soil salinity using AVIRIS-NG imaging spectroscopy and ensemble machine learning

• Published in: ISPRS journal of photogrammetry and remote sensing Vol. 200; pp. 191 - 212
• Main Authors: Das, Ayan, Bhattacharya, Bimal Kumar, Setia, Raj, Jayasree, G., Sankar Das, Bhabani
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2023
• Subjects: Ensemble machine learning; Hyperspectral remote sensing; interpretable Machine learning; Soil salinity index; EPO; DLensemble; GLM; COSRI; vij; vmax; AVIRIS-NG; NBDI; R1; R2; EC1:2; BDmax; R3; R4; GMM; Σi; EnMAP; μk; XGB; E1 to E4; RF; SDGs; RFensemble; BD; JAXA; DESIS; C; CNES; BI; IBD; SVR; BDR; SD; S; SNR; SI; ASI; Tmin; Hyperspectral remote sensing; SS; ESP; Soil salinity index; NASA-JPL; PRISMA; GBMensemble; CR; r; BNA; v; TIR; HEML; PLSR; HSSI; AWGN; CHI; di; Ci; HRS; DL; GPS; XRD; XRF; ANMB; EC; ML; DSM; HISUI; interpretable Machine learning; ei; MS; фk; NDSI; SHALOM; SWIR; dSm-1; MED; ANA; FWHM; Tmax; nm; ANN; ISRO; SAR; NS; VNIR; DLR; Tmean; VSlS; SRN; wi; GBM; Ensemble machine learning; SBG; nRMSEʹʹ and nRMSEʹ; BIC; PDP; HysPIRI; SSI; VStS; BRM; v^ij
• ISSN: 0924-2716; 1872-8235
• DOI: 10.1016/j.isprsjprs.2023.04.018

[Display omitted] Soil salinization is one of the major land degradation processes spread over millions of hectares of global land. Hyperspectral Remote Sensing (HRS) coupled with modern data mining approaches help in real-time and cost-effective assessment or monitoring of salt-affected soils. This study aimed at predicting soil salinity across five sites in India using the Airborne Visible-Infrared Imaging Spectrometer - Next Generation (AVIRIS-NG) data in low to moderately salt-affected cropland soils. We have identified four unique spectral absorption features having sensitivity towards soil salinity through a hybrid feature selection algorithm. Soil electrical conductivity (EC) was estimated using different machine learning (ML) based models such as random forest (RF), gradient boosting machines (GBM), and deep learning (DL). An ensemble of RF and DL models showed the best performance with the coefficient of determination (R2) of 0.89 and 0.55 and normalized root-mean-squared error of 0.15 and 0.16 in training and test datasets, respectively. We also proposed a new hyperspectral soil salinity index using Shannon entropy-based aggregation of selected absorption features. The newly proposed index outperformed other majorly used remote sensing-based salinity indices. It also showed a strong correlation with measured EC (r = 0.68) and ML-predicted soil EC (r = 0.78), both being significant at 1% level of significance. The index was effective in classifying HRS images into six distinct salinity classes. We also assessed the feasibility of applying the proposed salinity index for future hyperspectral missions through the simulation of various spectral-spatial resampling scenarios and estimated the optimal spectral and spatial resolution for salinity prediction. The hyperspectral salinity index can be directly estimated from HRS data without the need for time-consuming and expensive field samplings and used as a proxy to evaluate soil salinity status under field conditions.