Machine Learning Models for Regional Soil Monitoring and Global Data Prediction with Limited Data

Abstract

The application of machine learning (ML) in soil attribute mapping has significantly advanced over the past few decades. However, a key challenge remains: ML models are ‘data hungry’, and the limited number of soil observations poses a significant hurdle to their effectiveness. This thesis aims to use ML models trained on limited in-situ data to generate high-resolution maps of: (1) daily soil moisture (SM) over Tasmania (~80 m), and (2) global peat thickness and carbon stock estimation (~1 km).

Deep learning (DL) algorithms as a part of ML were utilized to develop models for mapping daily SM in Tasmania. We evaluated the models to estimate daily SM over time, using time series data from 39 in-situ measurement locations. To address the limited number of training dataset, the transfer learning model was used, using existing DL models trained with a much larger Australia dataset. These models incorporated inputs such as SMAP data, weather information, elevation maps, land use data, and multi-level soil properties maps to simultaneously generate daily SM estimates for surface and subsurface layers. The optimal models have been applied for automated mapping and integrated into the land monitoring system of Tasmania.

We further discussed the application of ML models for mapping global peat thickness and total carbon stocks. This emphasises the need for comprehensive global peatland data, which is currently missing from existing global soil map products. A global dataset of peat properties was compiled from available sources, combined with data augmentation techniques based on existing peat maps. The mapping process utilized a random forest algorithm was trained using 19 covariates, including geomorphological characteristics, climate variables, soil attributes, land cover, vegetation index, and remote sensing data, to generate maps at 1 km resolution. Predicted maps of peat thickness and carbon stock are made available to support modelling of peatlands across the globe.