| Abstract Submission No. | ABS-3.1-0072 |
| Title of Abstract | Coupling Geostatistics and Machine Learning for regionalized classification of land use and rock types from soil geochemical data |
| Authors | Abhishek Borah* |
| Organisation | Universidad de Chile |
| Address | Nabapur, Mazgaon Tezpur, Assam, India Pincode: 784001 E-mail: abhishekborah0@gmail.com |
| Country | India |
| Presentation | Oral |
| Abstract | This work deals with the regionalized classification of 10 lithological rock types and 10 land use/land cover (LULC) categories from the data of continuous features (concentration of 63 chemical elements and oxides in stream sediments from adjoining districts of Assam and Meghalaya) by means of a supervised ensemble learning algorithm, namely random forest. The novelty of our proposal is the deployment of a complementary set of features (�proxies�) at the sampled data points, calculated ingeniously by means of geostatistical techniques: (1) leave-one-out cross-validation (LOOCV), and (2) ordinary point cokriging with nugget-variance-filtering (for removing the short-scale spatial variation). Kriging or cokriging leverages the spatial correlation structure (variogram) but has limited capacity to incorporate the non-linear effects of environmental covariates (such as land use, lithology, etc.) on soil properties, which restricts high-precision prediction. In contrast, machine learning (ML) captures the quantitative relationship between soil properties and features derived from environmental covariates for predictive modelling, but cannot leverage the variogram, as it considers spatial data to be independent. This work highlights the importance of leveraging the strengths of geostatistical and ML approaches to ensure robust categorical data classification and spatial prediction of soil heavy metal content. We perform the cleaning and preparation of data of 4309 data points, provided by the Geological Survey of India (NGCM program) and accessible in the Bhukosh portal. We also added the LULC categories and the rock types to the data points. Compared to the traditional approach where rock types and LULC types are predicted from only the measured features, the random forest classifier that additionally uses the geostatistical proxies provides better performance scores (accuracy rate and Cohen�s kappa for measured features in combination with the geostatistical proxies are 92.61% and 0.9254 respectively, against the accuracy rate = 82.02% and Cohen�s kappa = 0.7854 for the traditional approach). |
| Keywords | Geostatistics, Machine Learning, Regionalized Data, LandUse/LandCover, Kriging |