Sensor Agnostic Semantic Segmentation of Structurally Diverse and Complex Forest Point Clouds Using Deep Learning

Forest inventories play an important role in enabling informed decisions to be made for the management and conservation of forest resources; however, process collecting inventory information is laborious. Despite advancements mapping technologies allowing forests digitized finer granularity than ever before, it still common measurements collected using simple tools such as calipers, measuring tapes, hypsometers. Dense understory vegetation complex structures can present substantial challenges point cloud processing tools, often leading erroneous measurements, making them less utility forests. To address this challenge, research demonstrates effective deep learning approach semantically segmenting high-resolution clouds from multiple different sensing systems diverse conditions. Seven datasets were manually segmented train evaluate model, resulting per-class segmentation accuracies Terrain: 95.92%, Vegetation: 96.02%, Coarse Woody Debris: 54.98%, Stem: 96.09%. By exploiting cloud, we also a method extracting Digital Terrain Model (DTM) clouds. This was applied set six that publicly available part benchmarking study DTM performance. The mean error 0.04 m relative reference with 99.9% completeness. These approaches serve useful steps toward fully automated reliable measurement extraction tool, agnostic technology used or complexity forest, provided has sufficient coverage accuracy. Ongoing work will see these models incorporated into tool structural metrics applications forestry, conservation, research.