Remote Sensing of Soil Physicochemical Properties with Sentinel-2A and ASD Field Spec-4 using Multiple Linear Regression Analysis and Artificial Neural Network

Abstract

In site-specific agriculture, mapping fine-scale spatial changes in soil characteristics are critical. Traditional soil analysis methods for mapping soil properties on a large scale are expensive, time-taking and not feasible. The current research looks at how remote sensing (RS) and geographic information system (GIS) approaches can be used to investigate the spatial variability of surface soil properties. An average of 51 samples were taken from tehsil Talagang of district Chakwal, Punjab, Pakistan. Lab analysis were carried out to determine soil texture and soil chemical properties. The objectives were to generate the predictive multiple linear regression modeling for soil physicochemical properties using Sentinel-2 and ASD Field Spec 4 data. The MLR has shown a significant (p<0.05) relationship of band-2, band-7, band-8 and band-8A with sand (%) (R2 = 0.19), OM (%) showed a significant (p<0.05) relationship with band-2, band-4, band-5 and band-11 with an R2 value of 0.23 and phosphorus obtained an R2 value of 0.22 while other properties did not show significant results. In the same way, OLS had obtained R2 value 0.11 for sand, K with R2 value of 0.19 and P had obtained R2 value of 0.10. Hence, both the techniques have not obtained significant results. Random Forest Regression has performed better than these for all soil properties. ASD data was modeled using SMLR and PLSR with sand properties. SMLR has sand (%) R2 value of 0.85, silt (%) of 0.71 and clay (%) 0.51 while with PLSR sand (%) has R2 value of 0.90, silt (%) 0.89 and clay (%) 0.83. Hence, PLSR performed better than SMLR in the case of hyperspectral data. Interpolation techniques were also used to predict soil properties, among which IDW has performed best. In conclusion, the DRS can be successfully used in the detection of soil properties as compared to multispectral data. Further, hyperspectral imagery can be used for most accurate results. Soil spectral libraries can be created using large amount of samples from the study area.