Machine learning based classification of peat layer thickness in Latvia using national forest inventory data

This study investigates the distribution and carbon content of organic soils in Latvia, leveraging machine learning techniques alongside remote sensing and National Forest Inventory (NFI) data to enhance the precision of organic soil mapping. Our approach integrates data from various sources, including airborne laser scanning (ALS) data, digital elevation models (DEM), depth-to-water (DTW) and wet area maps (WAM), and historical organic soil data. By classifying over 24,000 soil probing measurements across Latvia into distinct peat layer thickness categories, we develop a machine learning model that categorizes the thickness of the organic layer with notable accuracy. Our findings indicate that the model, particularly when employing the xgbTREE algorithm and over-sampling method, successfully identifies areas with peat layers thicker than 40 cm, demonstrating a significant improvement over traditional mapping methods. The study reveals an underestimation of organic soil coverage in Latvia by previous estimates, suggesting a broader distribution than recognized, with the model achieving an accuracy of 0.86 and a kappa value of 0.67. This research not only underscores the efficacy of integrating machine learning and remote sensing for soil mapping but also highlights the critical role of accurate data and models in determining organic soil distribution. The insights gained from this study are vital for policy-making and environmental planning, offering a more detailed understanding of Latviaʼs peatland resources and their conservation needs.