Evaluating machine learning algorithms to classify forest tree species through satellite imagery

Recent events indicate an uptick in forest fires in the western United States, prompting cities and organizations to develop a better understanding of forests and how to manage them. Tree species classification is important for forest management and carbon sequestration analysis. Currently, remote sensing stands as the prevalent method for the classification of tree species, land cover, etc. Researchers often use machine learning techniques for classification and general remote sensing. We hypothesized that it is possible to classify forest tree species with high classification accuracy using solely RGB values as the inputs for the machine learning models. We experimented with different machine learning algorithms such as Random Forest (RF), k-Nearest Neighbors (kNN), Gradient Boosting (GB), and Linear Discriminant Analysis (LDA) to classify forest tree species, specifically through multispectral Landsat 8 satellite imagery. Each algorithm was trained and validated using the same dataset and satellite imagery of the same region. Our findings indicated RF had the highest classification accuracy of 95.4% for validation on the same general region it trained on. kNN, GB, and LDA had classification accuracies of 81.6%, 76.4%, and 64.6%, respectively. Based on these results, we concluded that RF is the more accurate algorithm for classifying tree species through RGB satellite imagery. Our findings also indicate that model training and inference on the same general region result in higher classification accuracy. However, as the inference region changes, the classification accuracy reduces. In such cases, additional predictor variables, including trunk diameter, crown shape, and vegetation indices, could be introduced to improve classification accuracy.