Improving forest detection with machine learning in remote sensing data

Forest detection in remote sensing data is essential for important applications such as detection of area desertification, flooding simulation, forest health analysis, or conversion of digital elevation models. Existing techniques have open issues: they do not generalize well to different scenarios, they lack accuracy, and they require human intervention for input data characterization. To address these issues, in this work, we developed various classification models by using a variety of Machine Learning techniques, namely Convolutional Neural Networks (CNN), Random Forest ensembles (RF), and Support Vector Machines (SVM). Different CNN architectures were created specifically for the forest detection problem, and alternative feature extraction mechanisms were developed to support RF and SVM for this task. All these models were trained with SRTM and Landsat-8 satellite data, and their hyperparameters were optimized. Their effectiveness was assessed by using the Forest/No-Forest masks provided by JAXA as ground truth. Additionally, these models were compared against the JAXA's mask itself using expert-labeled data as ground truth. The experiments show promising results in terms of accuracy and generalization while presenting a reduced dependency on human intervention for characterizing data in both training and classification phases.