Validation of sentinel 2 based machine learning models for Czech National Forest Inventory

Abstract

The National Forest Inventory (NFI) of the Czech Republic provides essential data for forest management but requires significant time and resources. This study highlights the critical role of validating Sentinel-2-based machine learning models against real NFI data to ensure their reliability for forest monitoring. While satellite-based models offer a cost-effective alternative, their practical applicability depends on rigorous validation. We applied four commonly used machine learning models-Classification and Regression Trees, Random Forest, Support Vector Machine, and Naive Bayes-to Sentinel-2 imagery to estimate forest cover conditions. The Random Forest model achieved the highest overall accuracy (98.3 %). By systematically comparing model predictions with official NFI data, we address a key gap in remote sensing applications: the need for real-world validation beyond training datasets. Our findings demonstrate that properly validated Sentinel-2-based models can enhance large-scale forest monitoring, reducing the financial and labor burdens of traditional field surveys while ensuring data accuracy for sustainable forest management.