Tree Species Classification Using High-Resolution Satellite Imagery and Weakly Supervised Learning

Knowing vegetation type in an area is crucial for several applications, including ecology, land-use management, and infrastructure risk assessment. In combination with recent advancements in image processing, remote-sensing technology has been used to perform fast vegetation-type estimation and reduce the need for intensive and time-consuming field-based surveys. This article proposes a weakly supervised method based on deep learning to estimate tree species relying on multispectral high-resolution satellite images. We tested the approach against noisy labels, which often occur in real-world datasets. We validate our approach for a study area in Norway and Italy using images taken during different periods of the year. Our method significantly enhances the quality of the available forestry inventory dataset.