Optical Remote Sensing for Agroforestry Vegetation Segmentation & Biomass Estimation
Summary
Accurately estimating biomass in forestry systems is important for evaluating carbon offsets. Traditional methods of manual measurement are labor-intensive, costly, and use considerable resources. This research explores the potential of high-resolution optical remote sensing as a more cost-effective and easily available data source.
Current deep learning and machine learning techniques face challenges due to the limited availability of data and the complexity of labeling satellite imagery. Typically, these studies rely on high-resolution data (10m-30m), which may not be optimal for biomass monitoring, with very few exploring super-high-resolution satellite data.
Additionally, existing studies often employ either pixel-to-pixel segmentation or deep learning (DL) methodologies. Pixel-to-pixel approaches are limited as they fail to capture the surrounding of each datapoint. Meanwhile, DL methods are data-intensive, often impractical due to the scarcity of ground truth data and the laborious, sometimes inaccurate process of manual labeling.
This study proposes a novel vegetation segmentation technique that requires minimal labeling. We developed a texture-based machine learning segmentation approach which uses Local Binary Pattern and Random Forest, and which achieved up to 95\% accuracy on 4-classes segmentation. We also showed that this approach outperforms UNET, as well as that incorporating outcomes of this segmentation in the feature set significantly improves biomass estimation.