| 研究課題/領域番号 |
21K14261
|
|---|
| 研究種目 |
若手研究
|
| 配分区分 | 基金 |
|---|
| 審査区分 |
小区分22050:土木計画学および交通工学関連
|
|---|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
郭 直霊 東京大学, 空間情報科学研究センター, 客員研究員 (40897716)
|
|---|
| 研究期間 (年度) |
2021-04-01 – 2024-03-31
|
| 研究課題ステータス |
交付 (2022年度)
|
| 配分額 *注記 |
4,680千円 (直接経費: 3,600千円、間接経費: 1,080千円)
2023年度: 780千円 (直接経費: 600千円、間接経費: 180千円)
2022年度: 1,950千円 (直接経費: 1,500千円、間接経費: 450千円)
2021年度: 1,950千円 (直接経費: 1,500千円、間接経費: 450千円)
|
|---|
| キーワード | Post-disaster Monitoring / Deep Learning / Remote Sensing / Computer Vision / Change Detection |
|---|
| 研究開始時の研究の概要 |
This research addresses the challenge of achieving high performance recovery monitoring and related tasks based on inexact and inadequate training dataset. The weak-supervised learning, multi-task learning, transfer learning is proposed.
This research sets up an architecture which makes connecting between the research fields of urban mapping, change detection, and recovery monitoring.
Based on this research, more interdisciplinary datasets and methods are expected to for post-disaster recovery monitoring optimization to promote the efficiency and accuracy.
|
| 研究実績の概要 |
First, multi-temporal post-disaster remote sensing imagery and the ground-truth of changing has been collected.
Then, with the help of Multi-task Urban Mapping, the land cover semantic segmentation, object detection, and DSM in multi-temporal can be generated.
After that, an end-to-end deep learning model for change detection has been trained by data fusing and ensemble learning. Meanwhile, to solve the slight misalignment of multi-temporal imagery as well as the imbalanced land cover ratio, a specific loss function will be proposed.
Finally, the trained change detection model is capable of detecting land cover changes, such as the destroyed, new vacant, under construction, completed, etc.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
Progressing rather smoothly.
|
| 今後の研究の推進方策 |
Transfer Learning based Recovery Monitoring will be finished; in the next half year, some practical applications will be conducted.
For a large-scale and long-lasting post-disaster recovery monitoring, problems caused by multi-source remote sensing imagery such as the difference in land features, resolution, and spectrum, will be solved by transfer learning methods. Thus, a small amount of training datasets will be obtained for model transfer. After that, multi-temporal UAV aerial imagery will be put into the transferred multi-task urban mapping model for multi-temporal land feature extraction. Finally, the generated land features including semantic, object, and DEM will be utilized to achieve high-performance recovery monitoring via the transferred change detection model. In this year, the results of the aforementioned steps will be integrated, and the experimental post-disaster recovery monitoring will be deployed in cooperation with communication and companies. With the analysis of the experimental results, the rest of the time in FY2023 will be devoted to the conclusions of the theories, methods, and findings derived from this research. Additionally, the monitoring methods will be open for the society with practical applications, publications, and web service, which will lead to the future evolvement and application of post-disaster recovery monitoring in broader usages.
|
