Hyderabad: An ISRO-National Remote Sensing Centre (NRSC) study has combined satellite imagery, field-tested soil samples and machine learning to map soil health across farmland in Indukurpet mandal of SPSR Nellore district in Andhra Pradesh.The researchers used Landsat-8 satellite data to predict soil pH and electrical conductivity, or EC, an indicator of salinity. The method produced digital maps showing local variations in soil acidity and salinity, including smaller patches that could be missed by widely spaced field sampling.The study, titled “Machine learning-based prediction of soil pH and EC using Landsat-8 images”, was published online in Environmental Earth Sciences on June 15, 2026.175 soil samples collectedThe study covered about 144 sq km in Indukurpet mandal, which has an average elevation of four metres and a hot, humid climate. Paddy is a major crop in the region, along with legumes.The researchers had collected 175 surface-soil samples from 15 village locations between Feb 12 and 27, 2025, using stratified random sampling. Landsat-8 imagery from the same month was used to maintain consistency between the field and satellite observations.The laboratory-tested samples were combined with six Landsat-8 spectral bands, ranging from blue to short-wave infrared, and 12 spectral indices. These included salinity indices such as the Normalised Difference Salinity Index and Canopy Response Salinity Index, as well as vegetation indices such as the Normalised Difference Vegetation Index and Soil-Adjusted Vegetation Index. The satellite data were processed through Google Earth Engine.Short-wave infrared bands and salinity indices contributed more to the predictions than vegetation indices, indicating that the models obtained more useful information from soil moisture and mineral-related spectral responses than from plant greenness alone.Maps can guide field testingThe researchers also generated spatial prediction and uncertainty maps. While the prediction maps identified variations in acidity and salinity, the uncertainty maps showed areas where the models were less certain and where further field sampling may be required.The approach could help agricultural departments concentrate soil testing in high-risk or uncertain zones instead of carrying out intensive sampling across an entire region. The maps could also support decisions on crop suitability and fertiliser application.The findings are relevant to crops such as paddy, which is sensitive to salinity and requires a soil pH of about 6.5 to 7.5 and electrical conductivity below 2 dS/m for optimum yields.The researchers combined remote sensing with machine learning, offering a cost-effective alternative to labour-intensive field surveys and laboratory analysis, which have limited spatial coverage. The method could support soil-health monitoring, sustainable farming and land-use planning over large agricultural areas.Model performanceAmong the five machine learning models tested, Random Forest (RF) achieved the highest accuracy for predicting EC, while Gradient Boosting Regression (GBR) performed best for predicting soil pH.The research team comprised Suneetha C and Lakshmi Sutha Kumar of the National Institute of Technology Puducherry; Sreenivas K and Tarik Mitran of NRSC, ISRO, Hyderabad; and Venkatachalam K of Audisankara University– Among the five machine learning models tested, Random Forest (RF) achieved the highest accuracy for predicting EC, while Gradient Boosting Regression (GBR) performed best for predicting soil pH.The research team comprised Suneetha C and Lakshmi Sutha Kumar of the National Institute of Technology Puducherry; Sreenivas K and Tarik Mitran of NRSC, ISRO, Hyderabad; and Venkatachalam K of Audisankara University.
