Title | Leveraging deep neural networks to map caribou lichen in high-resolution satellite images based on a small-scale, noisy UAV-derived map |
| |
Author | Jozdani, S; Chen, D; Chen, W ; Leblanc, S G ; Prévost, C; Lovitt, J ; He, L ; Johnson, B A |
Source | Remote Sensing vol. 13, issue 14, 2658, 2021 p. 1-24, https://doi.org/10.3390/rs13142658 Open Access |
Image |  |
Year | 2021 |
Alt Series | Natural Resources Canada, Contribution Series 20210345 |
Publisher | MDPI AG |
Document | serial |
Lang. | English |
Media | paper; on-line; digital |
File format | pdf; html |
Province | Quebec |
Area | Manicouagan Reservoir; Canada |
Lat/Long WENS | -68.9000 -68.7167 50.6250 50.5833 |
Subjects | Science and Technology; remote sensing; unmanned aerial vehicles |
Illustrations | satellite imagery; location maps; diagrams; tables; graphs |
Program | Canada Centre for Remote Sensing Remote Sensing Science Program - Optical methods and applications |
Released | 2021 07 06 |
Abstract | Lichen is an important food source for caribou in Canada. Lichen mapping using remote sensing (RS) images could be a challenging task, however, as lichens generally appear in unevenly distributed, small
patches, and could resemble surficial features. Moreover, collecting lichen labeled data (reference data) is expensive, which restricts the application of many robust supervised classification models that generally demand a large quantity of labeled
data. The goal of this study was to investigate the potential of using a very-high-spatial resolution (1-cm) lichen map of a small sample site (e.g., generated based on a single UAV scene and using field data) to train a subsequent classifier to map
caribou lichen over a much larger area (~0.04 km2 vs. ~195 km2) and a lower spatial resolution image (in this case, a 50-cm WorldView-2 image). The limited labeled data from the sample site were also partially noisy due to spatial and temporal
mismatching issues. For this, we deployed a recently proposed Teacher-Student semi-supervised learning (SSL) approach (based on U-Net and U-Net++ networks) involving unlabeled data to assist with improving the model performance. Our experiments
showed that it was possible to scale-up the UAV-derived lichen map to the WorldView-2 scale with reasonable accuracy (overall accuracy of 85.28% and F1-socre of 84.38%) without collecting any samples directly in the WorldView-2 scene. We also found
that our noisy labels were partially beneficial to the SSL robustness because they improved the false positive rate compared to the use of a cleaner training set directly collected within the same area in the WorldView-2 image. As a result, this
research opens new insights into how current very high-resolution, small-scale caribou lichen maps can be used for generating more accurate large-scale caribou lichen maps from high-resolution satellite imagery. |
Summary | (Plain Language Summary, not published) Lichen is an important food source for caribou in Canada. Lichen mapping using remote sensing images is a challenging task, however, as lichens generally
appear in unevenly distributed, small patches, and could resemble surficial features. Moreover, collecting lichen ground truth data is expensive, which restricts the application of many robust supervised classification models that generally demand a
large quantity of labeled data. The goal of this study was to investigate the potential of using a very-high-spatial resolution (1-cm) lichen map of a small sample site (e.g., generated based on a single UAV scene and using field reference data) to
train a subsequent classifier to map caribou lichen over a much larger area (~0.04 km2 vs ~195 km2) and a lower spatial resolution image (in this case, a 50-cm WorldView-2 image). We used a Teacher-Student semi-supervised learning approach that
trains Teacher and Student networks iteratively on both ground-truthed and not-ground-truthed data. This approach produced lichen maps at the WorldView scale with an ~85% overall accuracy. |
GEOSCAN ID | 329023 |
|
|