Spatial mapping of predicted soil orgainc carbon in Thailand : a machine learning approach | |
| Author | Nitiwat Wongchanla |
| Call Number | AIT Thesis no.AE-25-01 |
| Subject(s) | Soil mapping Soils--Carbon content--Remote sensing Machine learning |
| Note | A Thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Agricultural Systems and Engineering |
| Publisher | Asian Institute of Technology |
| Abstract | Soil organic carbon (SOC), functioning as a significant factor of soil productivity and health, supports agricultural productivity and environmental sustainability. Therefore, estimating SOC content is crucial to assess carbon sequestration and evaluate the effectiveness of land management practices. The primary objective of this research was to construct a machine learning (ML) approach to predict how SOC is distributed spatially across Udon Thani landscapes. Ninety-six distinct environmental variables including remotely sensed, terrain, and climatic attributes were derived from Landsat 8 imagery, digital elevation model (DEM) data, and WorldClim data respectively to develop the best SOC prediction model using random forest (RF), artificial neural networks (ANN), Cubist, and boosted regression trees (BRT) algorithms. To avoid overfitting predictors during model development, statistically significant environmental variables were identified using Spearman’s rank correlation and Student’s t-test at p < 0.05. The analysis demonstrated that the contribution of environmental variables was not consistent among the various ML models. The key variables included Channel Network Base Level (CNBL) (22.04%), Elevation (21.45%), Valley Depth (16.32%), Standardized Height (14.07%), and Topographic Wetness Index (12.27%) for the ANN model. Standardized Height (34.69%), Valley Depth (33.48%), and Relative Slope Position (22.48%) were most significant for the BRT model. The Cubist model was predominantly influenced by CNBL (84.88%) whereas the RF model identified Modified Catchment Area (12.12%), Elevation (11.95%), Valley Depth (11.43%), Slope Length (11.06%), CNBL (11.06%), and Standardized Height (10.39%) as the dominant predictors. In comparison to other ML approaches, the results highlighted that the ANN model exhibited the best performance in predicting the spatial variability of SOC, achieving an R² of 0.64, an RMSE of 0.06%, and an MAE of 0.05%. The study demonstrates complex interactions between SOC and terrain attributes in conjunction with climatic factors and emphasizes the need for location-specific management strategies to sustain agricultural productivity. Future research should aim to improve model accuracy by incorporating additional soil physicochemical properties. |
| Year | 2025 |
| Type | Thesis |
| School | School of Environment, Resources, and Development |
| Department | Department of Food, Agriculture and Natural Resources (Former title: Department of Food Agriculture, and BioResources (DFAB)) |
| Academic Program/FoS | Agricultural Systems and Engineering (ASE) |
| Chairperson(s) | Himanshu, Sushil Kumar |
| Examination Committee(s) | Singh, Jai Govind;Datta, Avishek |
| Scholarship Donor(s) | The Royal Thai Government Fellowship |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2025 |