Monitoring salinity in bare soil based on Sentinel-1/2 image fusion and machine learning

• Published in: Infrared physics & technology Vol. 131; p. 104656
• Main Authors: He, Yujie, Zhang, Zhitao, Xiang, Ru, Ding, Binbin, Du, Ruiqi, Yin, Haoyuan, Chen, Yinwen, Ba, Yalan
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2023
• Subjects: GS(Gram-Schmidt); Image fusion; Multispectral; Soil salinity; Synthetic aperture radar; Multispectral; GS(Gram-Schmidt); Soil salinity; Image fusion; Synthetic aperture radar
• ISSN: 1350-4495; 1879-0275
• DOI: 10.1016/j.infrared.2023.104656

•Radar image and spectral images were fused for monitoring soil salinity.•Soil salinity inversion models were built based on Sentinel-1/2 fusion and machine learning.•GS was used to fuse images of Sentinel-1 and Sentinel-2. In the application of satellite remote sensing data to salinity inversion in bare soil, multispectral satellites enjoy richer wavelength bands compared with synthetic aperture radar (SAR), but their vulnerability to cloudy weather limits the inversion accuracy. In contrast, SAR has stronger penetration, is less affected by cloudy weather, and can work around the clock. Therefore, in order to monitor soil salinity over a longer period of time and under various climatic conditions, and to further improve the accuracy of salinity inversion, we fused Sentinel-1 and Sentinel-2 images and investigated the feasibility of SAR and multispectral satellite image fusion in salinity inversion in this paper. And we conducted experiments in Hetao Irrigation Area, Inner Mongolia, China. Gram-Schmidt(GS) was used to fuse Sentinel-1 and Sentinel-2 images at the same time, and maximum-normalization was carried out on the DN value of the fused image. Then some remote indices conducted with the normalized DN, together with high relevant fused single bands were input into Back Propagation(BP), Support Vector Machine(SVM), Random Forest(RF) as independent variables after two-tailed significance test and Variable importance in projection(VIP) selection, for soil salinity inversion model construction. Finally, the inversion results of the study area were evaluated with R2, RMSE, RPD, RSS, and RPIQ. The results of this study showed that the bare soil salinity inversion based on the fusion of VV(Sentinel-1) and Sentinel-2 had a higher accuracy than that on the fusion of VH (Sentinel-1) and Sentinel-2 image and that without the fusion. Among the three machine learning models, random forest (RF) achieved the most satisfying results (R2 = 0.801, RMSE = 0.686), followed by SVM(R2 = 0.624, RMSE = 0.875) and BP(R2 = 0.613, RMSE = 0.918). This study demonstrated the feasibility of fusing Sentinel-1 radar images with Sentinel-2 multispectral images to improve the accuracy of soil salinity inversion, and constructed the soil salinity inversion model based on machine learning and GS fusion.