講演詳細

経営部門[Forest Management]

日付	2024年3月8日
開始時刻	ポスター発表
会場名	531
講演番号	PD-34（学生ポスター賞審査対象）
発表題目	Combining Graph and Convolutional Neural Networks with multi-sensor remote sensing for forest type classification Combining Graph and Convolutional Neural Networks with multi-sensor remote sensing for forest type classification
所属	The University of Tokyo
要旨本文	Forest type classification is fundamental and has potential implications for environmental protection and climate change mitigation. Yet, there are difficulties in forest type classification such as multi-scale features, heterogeneous boundary, mountainous terrains and unbalanced datasets. We introduced multi-source remote sensing (RS) dataset by transferring learning on large natural dataset (RGB bands) of ImageNet 22K, a region segmentation and gated graph convolution module network (GGCM). Extensive experiments were conducted in two study sites in central and northern Japan with different tree species composition and management strategies for the purpose of cross-regional forest type classification task. Results indicated that multi-sensor RS data yielded superior classification accuracy than the single sensor RS datasets. Pretrained weights improved accuracy greatly, and GGCM was capable of retaining multi-scale and spatially topological features well in both research sites.
著者氏名	○Pei Huiqing1 ・ Owari, Toshiaki1,2 ・ Tsuyuki, Satoshi1 ・ Hiroshima, Takuya1
著者所属	1Department of Global Agricultural Sciences, The University of Tokyo ・ 2The University of Tokyo Hokkaido Forest, The University of Tokyo
キーワード	deep learning, gated graph convolution module network, multi-sensor remote sensing, forest type classification
Key word	deep learning, gated graph convolution module network, multi-sensor remote sensing, forest type classification