Tree Species Mapping on Sentinel-2 Satellite Imagery with Weakly Supervised Classification and Object-Wise Sampling

Abstract

Information on forest composition, specifically tree types and their distribution, aids in timber stock calculation and can help to better understand the biodiversity in a particular region. Automatic satellite imagery analysis can significantly accelerate the process of tree type classification, which is traditionally carried out by ground-based observation. Although computer vision methods have proven their efficiency in remote sensing tasks, specific challenges arise in forestry applications. The forest inventory data often contain the tree type composition but do not describe their spatial distribution within each individual stand. Therefore, some pixels can be assigned a wrong label in the semantic segmentation task if we consider each stand to be homogeneously populated by its dominant species. Another challenge is the spatial distribution of individual stands within the study area. Classes are usually imbalanced and distributed nonuniformly that makes sampling choice more critical. This study aims to enhance tree species classification based on a neural network approach providing automatic markup adjustment and improving sampling technique. For forest species markup adjustment, we propose using a weakly supervised learning approach based on the knowledge of dominant species content within each stand. We also propose substituting the commonly used CNN sampling approach with the object-wise one to reduce the effect of the spatial distribution of forest stands. We consider four species commonly found in Russian boreal forests: birch, aspen, pine, and spruce. We use imagery from the Sentinel-2 satellite, which has multiple bands (in the visible and infrared spectra) and a spatial resolution of up to 10 meters. A data set of images for Leningrad Oblast of Russia is used to assess the methods. We demonstrate how to modify the training strategy to outperform a basic CNN approach from F1-score 0.68 to 0.76. This approach is promising for future studies to obtain more specific information about stands composition even using incomplete data.