| 研究実績の概要 |
First, multi-source remote sensing imagery with related ground truth, including semantic segmentation annotation, object bounding box, and DEM are collected.
Then, weak supervision based multi-task deep learning methods are formulated to train the urban mapping model, which can conduct semantic segmentation, object detection, and DEM generalization simultaneously even with the inexact and inadequate training dataset.
Meanwhile, a learnable weighting method for multi-loss functions combination is proposed to achieve automatic fine-tuning.
The multi-task urban mapping will serve the following steps with a high-performance multi-task mapping model, and provide additional land cover mapping results as well.
|
| 今後の研究の推進方策 |
Post-disaster Change Detection will be the finished in FY2022.
First, multi-temporal post-disaster remote sensing imagery and the ground-truth of changing will be collected. Then, with the help of (A) 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 will be 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 will be capable of detecting land cover changes, such as the destroyed, new vacant, under construction, completed, etc.
|
