FOREST-RELATED SDG ISSUES MONITORING FOR DATA-SCARCE REGIONS EMPLOYING MACHINE LEARNING AND REMOTE SENSING - A CASE STUDY FOR ENA CITY, JAPAN

We proposed a combined machine learning approach with a deep convolutional neural network (CNN) to monitor forest utilization toward Sustainable Development Goals (SDGs) for data-scarce regions. First, we employed the Random Forest (RF) classifier using Google Earth Engine (GEE) for forest mapping. Then, we designed a deep CNN architecture that works for tree species/age mapping from coarse and polygonal ground-truth data. The proposed network has U-shape and comprises 3D Atrous Convolutions. The model was optimized by a weighted cross-entropy loss function. We trained the model with times-series Sentinel 1, 2, and Digital Elevation Model (DEM) data with sparse annotations. Our proposed models achieved 94.5% overall accuracy (OA) for forest mapping, 77.80% (OA) for tree species, and 81.74% (OA) for tree age classification, respectively in Ena city, Japan. The outcome of our study indicates the potential of remote sensing and machine learning in monitoring forest development, conservation, and utilization toward SDGs from coarse ground-truth data. Our source code for the implementation is available at: https://github.com/anhp95/forest_attr_segment