2020 Fiscal Year Research-status Report
Deep forest for disaster monitoring with multi-temporal and multi-modal earth observation data
| Project/Area Number |
19K20309
|
|---|
| Research Institution | Institute of Physical and Chemical Research |
|---|
Principal Investigator |
Xia Junshi 国立研究開発法人理化学研究所, 革新知能統合研究センター, 研究員 (00830168)
|
|---|
| Project Period (FY) |
2019-04-01 – 2022-03-31
|
| Keywords | Damage mapping / Machine learning / Deep forest / Multi-source datasets / Siamese networks |
|---|
| Outline of Annual Research Achievements |
The research achievements can be summarized as follows:
1. We have made several damage responses with Group Earth Observation (GEO), such as 2020 Chinese Summer Flood, 2020 Beirut explosion, 2021 Indonesia Floods and landslides. In the disaster responses, we have applied the image preprocessing (i.e., noise filtering), transfer learning, Siamese convolutional neural networks (CNN), noisy label learning and post-processing to provide the change detection and building damage mapping.
2. We have constructed very high-resolution (1m) Gaofen-3 Synthetic Aperture Radar(SAR) datasets for building semantic segmentation. For the datasets, we compare the performance of difference baselines, and give the guidelines and roadmap for the future studeis. The datasets will be extend for the use of damage mapping.
3. We have constructed the multimodal (1 m high-resolution optical and SAR) datasets (more than 10 events) for damage mapping and proposed a general framework of learning from multimodal and multitemporal earth observation data for building damage mapping.
4. We have developed the ensemble of diverse Siamese CNN, such as the Unet with different encoders and attention mechanism, for building damage mapping.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The project was performed smoothly. Based on the large scale building damage datasets, we have developed multiple methods to improve the accuracy, and provide the building damage mapping for rapid damage response.
|
| Strategy for Future Research Activity |
We have already developed the methods for the xView2 and our multiple source datasets. In the next fiscal year, my research plan includes
1. Develop the efficient building damage mapping with the combination of deep forest and its variants
2. Enlarge the xView2 database with the deep learning models and Maxar open disaster program
3. Develop the transfer learning to predict the building damage mapping of new events from the previous events
|
| Causes of Carryover |
We will consider to submit the papers to several international conference and journals. We will buy the high-resolution optical and SAR datasets.
|
