Mapping forest felling activities in Latvia from Sentinel-2 satellite imagery using machine learning

Sammanfattning

This thesis develops and evaluates a Random Forest (RF) machine learning classification framework for detecting clear-cut forest felling events using Sentinel-2 multispectral satellite data. While RF classification is widely applied in forest monitoring, its performance for operational clear-cut detection in Latvian semi-boreal forests remains insufficiently studied. This study aims to assess the applicability and robustness of RF model across multiple forest districts using multi-temporal satellite data. Model training was conducted using four study areas within Latvian State Forests. For each area, cloud-masked Sentinel-2 imagery representing pre-felling (2018) and post-felling (2024) conditions was selected, resulting in multi-temporal datasets. Median spectral reflectance features were derived based on historical felling records. Model evaluation included two approaches: internal validation within the training areas (self-testing), and external validation using a fifth forest district not included in training. Performance was assessed using overall accuracy, user accuracy (precision), producer accuracy (recall), and F1 score. The results demonstrate that the RF model achieved over 70% user and producer accuracies in the unseen test area, with overall accuracy exceeding 95%. Analysis of feature importance showed that visible and shortwave infrared spectral bands contributed most significantly to distinguishing felled and non-felled areas. The inclusion of vegetation indices improved classification accuracy by 1–4%. Misclassifications were primarily related to sanitary fellings and cloud-contaminated pixels. The findings indicate that Random Forest classification applied to Sentinel-2 data is capable of detecting clear-cut forest felling events in Latvian forests. The study highlights the importance of spectral band selection, vegetation indices, and robust validation strategies, contributing to improved forest monitoring capabilities in semi-boreal environments.