Assessment of temporal datasets constructed with WFI imagery for mapping burned areas

This study evaluates the performance of binary classifications of burned areas using satellite imagery time series provided by the WFI sensors on board the CBERS-4A and AMAZONIA-1 satellites. Five machine learning algorithms were applied for supervised classification: Support Vector Machine (SVM), Random Forest (RF), XGBoost, Simple Recurrent Neural Network (SimpleRNN), and Long Short-Term Memory (LSTM). The models were trained and validated on 2020 data and tested on unseen data (dataset from 2021). Performance was assessed using precision, recall, and F1-score metrics on 30\% validation subsets, with the Intersection over Union (IoU) applied to the full test datasets. The results indicated superior model performances on the 2020 dataset due to the inclusion of seen data. However, on the 2021 dataset, SimpleRNN and LSTM underestimated burns, while SVM, RF, and XGBoost overestimated them.