Deforestation Monitoring Using Machine Learning Methods and Time-Series Satellite Data

"This thesis investigates the application of machine learning methods to the task of deforestation monitoring using time-series satellite data. The objective is to assess how different algorithmic approaches perform in detecting forest loss based on spectral and vegetation index signals derived from multi-temporal optical imagery. A comparative framework was developed to benchmark traditional classifiers and deep learning architectures with respect to their accuracy, computational efficiency, and interpretability.

The methodology combines pixel-level vegetation time series with stratified training samples and evaluates model outputs against validated reference data. Models were trained and tested in a cloud-based environment using consistent preprocessing and feature extraction pipelines. Key evaluation metrics were used to characterize the strengths and limitations of each approach.

The results show that, under the right conditions, well-optimized traditional machine learning models can achieve deforestation detection performance comparable to that of deep learning techniques. This highlights the importance of careful feature engineering and the quality of ground truth labels. While recurrent neural networks excel in capturing complex temporal dynamics, they come with substantial computational costs and implementation complexity. In contrast, classical models such as ensemble methods or linear classifiers offer competitive performance when paired with informative input representations and are better suited for scalable or resource-constrained monitoring systems.

These findings contribute to the broader discussion on operational deforestation monitoring by demonstrating that model choice must be aligned with the intended use case—whether focused on early-warning alerts, policy reporting, or high-throughput analysis—and by identifying practical trade-offs between accuracy, explainability, and computational demand."