Understanding Remote Sensing Images
The automatic extraction of map information, such as land cover labels and elevation models from remote sensing imagery, is our core project. Machine learning is a powerful tool to automate mapping, but the quality and quantity of training data are the key. We build world-leading benchmark data for land cover mapping addressing geographic equity [Xia, WACV'23]. We also work on mapping from incomplete training data [Li, JSTARS'22]. How to recognize changes from images of different modalities and automatically update maps is also a key challenge [Chen, TGRS'22].
Solving Inverse Image Problems
We work on solving image inverse problems, such as compressive spectral imaging, super-resolution, denoising, and inpainting to obtain high-spatio-spectral-temporal-resolution images that exceed hardware limitations. We develop algorithms based on deep learning and matrix and tensor decomposition that contribute to the generation of higher-order products from airborne and satellite data. Our recent work include cross-modal super-resolutioin [Dong, ECCV'22] and neural architecture search for compressive spectral imaging [He, ECCV'22].
Toward a Sustainable Future
Our research is grounded in solving global issues toward, such as climate change, large-scale natural disasters, and food problems. Our recnet work includes 3D change (e.g., inudation depth and terrain deformation) estimation during disasters [Yokoya, TGRS'22], biomass and carbon stocks estimation in mangroves [Le, IJRS'21], and crop type classification [Xia, IJAEOG'23], in collaboration with related institutions and researchers in Japan and overseas. In real-world problems, the lack of training data is a more serious issue. A unique feature of our approach is the successful use of heterogeneous data from different disciplines as training data.