This study focuses on developing an ensemble learning model to detect sugarcane burning in the northern and central coastal regions of Peru during the period 2017-2022. Burning sugarcane residues is a widespread practice
that has significant negative effects on soil health, the environment, and public health. The proposed model aims to accurately identify burning events by leveraging satellite imagery and advanced machine learning algorithms.
Specifically, the study integrates deep learning architectures like U-Net and tree-based classifiers such as Gradient Boosting Decision Trees (GBDT), combining their strengths for precise classification of burned areas. The dataset
for training the model was specifically curated, including both multispectral satellite data and ground-truth labeling.
The results showed that the ensemble model outperformed individual classifiers in terms of metrics like F1-score (90.6 %), Kappa index (87.5 %), and IoU (82.8 %), providing a scalable solution applicable to other agricultural contexts. The findings have significant
implications for environmental monitoring and regulatory compliance, supporting authorities such as OEFA in identifying and mitigating the impact of sugarcane burning practices.
