Comparing Multi-Training and Multi-Projection Approaches in Random Forest Classification for Earth Observation Land Cover Analysis

This study investigates whether, when using multispectral imagery, it is preferable to retrain a random forest model once a year or to train it for a single year and then use that model to predict all subsequent years. We utilized Sentinel-2 images as the primary input features along with environmental data for the study. The targets representing both natural and semi-natural habitats for classification were used in an area of central Italy. By excluding all burned areas since 2001, we assessed the area’s stability over the past nine years. The approach involved conducting a twin experiment to compare the results of model training and predictions under two distinct scenarios. In the initial experiment, named SINGLE, the model was trained using data from one year (2023) and subsequently applied to generate land cover maps spanning from 2016 to 2024. In the second experiment, known as MULTI, the model undergoes annual retraining before forecasting the land cover for each respective year. The findings reveal that in the MULTI experiment, the accuracies are not only superior to those in the SINGLE experiment but also exhibit consistent stability, highlighting the effectiveness of annual model retraining.