Authors– Suresh Kumar and Justin George K

In recent times machine learning (ML) models have emerged as potential tool in Digital Soil Mapping (DSM). It is an innovative method that generates precise soil information employing geospatial technology, including remote sensing, geographic information systems (GIS), geographic position system (GPS) and statistical modelling. Environmental covariates derived using RS and GIS are integrated with soil data employing ML models facilitating to generate soil information. Digital soil mapping activities has largely expanded across the world since the early 2000s. In the geospatial era, DSM has emerged as the potential method in monitoring of soil resources as well as empowering to digital agriculture promoting informed decision making to planners, user agencies and farmers. Ministry of Agriculture Farmers Welfare, Govt. of India initiated the National Soil Mapping Programme under Digital Agriculture Mission in year 2025.

Spatial variability of soils in the landscape are controlled by soil forming factors defined by Jenny (1941) as S=f (cl, o, r, p, t..) where, soil type variation is primarily accounted by parent material (p) (geomorphology / geology), conditioned by relief (r) and organism (o) (land use /cover) in a climate (cl) region over a period of time (t). In conventional method of soil mapping, polygons as soil map units were delineated following Jenny’s clorpt model. Soils in these map units are characterized by examining pedons (soil profiles).

Soil information with precise geographic location in association with environmental covariates enabled soil surveyors or pedologist in realization of digital soil mapping techniques for mapping of soil properties on pixel basis. Knowing the geographic location of soil properties over the region enabled to establish / explore their spatial relationship with environment covariates in a systematic way employing machine-learning models. McBratney (2003) explored this relationship and proposed SCORPAN Model as S=f (S, C, O, R, P, A, N) which forms the basis of Digital Soil Mapping method. In this model, s: soil attributes at a point is expressed as a function of s: legacy soil information of the area, c: climate properties, o: organisms (including land cover and natural vegetation); r: topography (terrain attributes), p: parent material (including lithology), a: age (time factor); n: space (geographic position).

The SCORPAN model acts as guide to generate various environmental covariables. The SCORPAN model is almost similar to clorpt model except geographic location (n) of soil data. Today availability of various open-source Earth Observation (EO) data (Copernicus, Landsat, Sentinel 1 & 2 and IRS Resourcesat 1-2 data), WorldClim climate database products etc. offered by several agencies assist us in generating these environmental covariates. Besides these, geospatial analysis platforms such as Google Earth Engine (GEE) is offering seamless long-term high spatial and temporal resolution remote sensing data to generate soil, geology and vegetation related spectral indices of varying temporal period. Digital Elevation Models (DEM) used to derive primary and secondary terrain variables form another most important source of environment covariates. Climate data (rainfall, temperature, evapotranspiration etc.) available at various resolutions as spatial gridded data as well as vegetation data products offered by BHUVAN, VEDAS and NASA EO websites also form prominent environment covariables. ML models such as Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Regression (SVR) are commonly being used and are available within statistical software packages, such as “CART”, “RF” and “Ranger,” of R software.  Several websites also offer ML algorithms / models to implement online DSM. Among the various ML Techniques, Random Forest (RF) emerged as most versatile model of DSM. Among various soil properties, soil organic carbon (SOC), sand, silt clay, pH, EC, soil nutrients (N, P &K) are successfully mapped with DSM, across the globe.

DSM has ushered a new era of generating updated soil information required at various scales to meet the increasing needs of diverse users for distinct applications related to digital agriculture, nutrient management, soil carbon dynamics, soil quality assessment etc. It enables the monitoring of land degradation, desertification and soil quality / health, as well as the quantification of soil carbon sequestration potential to mitigate climate change impacts, while supporting the evaluation of ecosystem services, environmental sustainability and food security.  DSM provides a valuable insights of dynamic soil properties and is more efficient than conventional soil mapping (CSM) but not the complete replacement of CSM.

A case study: Das et al. (2025) used various machine learning (ML) models and compared their predictive ability for digital mapping of soil properties in a hilly watershed located in the Indian Himalayan region in Dehradun district, Uttarakhand, India. Prediction performances of different ML algorithms namely, multivariate adaptive regression splines (MARS), support vector regression (SVR), and artificial neural network (ANN) were compared with multiple linear regression (MLR) models and further employed for mapping soil properties.

References

• Das, P., Kumar, S., Justin, K and Ahmad, T.  (2024). A Comparative Study on the Predictive Ability of Machine Learning Techniques for Spatial Mapping of Soil Properties in Indian Himalayan Region. Available at http://dx.doi.org/10.2139/ssrn.4658128
• Jenny, H., 1941. Factors of Soil Formation: A System of Quantitative Pedology. McGraw-Hill Book Company, Inc.
• McBratney, A.B., Mendonca, M.L., Minasny, B. (2003). On digital soil mapping, Geoderma,117 (1-2): 5-32.