10:45 AM - 12:15 PM
[ACG42-P05] Continuously river discharge observation using deep learning-based semantic segmentation
Keywords:River flow, Hydrological observation, Flow observation, Still image, Image analysis, Machine Learning
The acquisition of river discharge data is the most important task for river management. However, observing the river discharge safely under various weather conditions or surrounding environments isn't easy. Existing methods usually multiply cross-sectional area by flow velocity that is observed by current meter. Kudo et al. (2022) developed a new observation method to estimate the river discharge using only one still image. The benefits of that can be concluded as follows. First, as only one still image is needed, the data capacity of our proposed method is very low. Second, there is less limitation to the location of the camera as if the river can be shot. Kudo et al. (2022) extracted water body from one still image by machine learning –based segmentation method, Trainable Weka Segmentation (Arganda-Carreras et al., 2017), and estimated the width of the river and the area of water using the mask of the extracted water body. Then the river discharge was calculated by the correlation between the river discharge and the width of the river and the area of the water body.
However, the prediction results of traditional machine learning-based segmentation methods can be unpleasant when the input images are tricky, for example, images with shadow or blown-out highlights. Therefore, we propose to solve this problem by using deep learning-based semantic segmentation methods to extract water body from still image. To be specific, in this paper we produced an original dataset that includes over 2,000 images of various rivers. We trained 3 SoTa semantic segmentation models using our original dataset and the open source dataset, Atlantis (Erfani et al., 2022) and conducted comprehensive experiments.
We estimated the width of a river and the area of the water body using DeepLabV3 Plus (Chen et al., 2018), PointRend (Kirillov et al., 2020), and SegFormer (Xie et al., 2021). The experiment used images which were taken at an experiment site located in a natural river, where a water level gauge was set, allowing us to get the ground truth of the river discharge. We compared the pixel number of the width of the river and the area of the water body to river discharge ground truth and found a strong correlation between them, similar to the result of Kudo et al. (2022). There were no large differences in the value of the river discharge between the three deep learning models. Among them, SegFormer has the highest accuracy in predicting the water body. We concluded that the performance of these models is sufficient enough to estimate the river discharge.
On the other hand, some tricky images, such as low angle of incidence of sunlight (images taken at dawn or early evening), riverbeds strewn with small stones, or images with large shadows, are still difficult to be predicted accurately. We plan to add these tricky images to our dataset and perform some image preprocessing before the prediction in the future.
However, the prediction results of traditional machine learning-based segmentation methods can be unpleasant when the input images are tricky, for example, images with shadow or blown-out highlights. Therefore, we propose to solve this problem by using deep learning-based semantic segmentation methods to extract water body from still image. To be specific, in this paper we produced an original dataset that includes over 2,000 images of various rivers. We trained 3 SoTa semantic segmentation models using our original dataset and the open source dataset, Atlantis (Erfani et al., 2022) and conducted comprehensive experiments.
We estimated the width of a river and the area of the water body using DeepLabV3 Plus (Chen et al., 2018), PointRend (Kirillov et al., 2020), and SegFormer (Xie et al., 2021). The experiment used images which were taken at an experiment site located in a natural river, where a water level gauge was set, allowing us to get the ground truth of the river discharge. We compared the pixel number of the width of the river and the area of the water body to river discharge ground truth and found a strong correlation between them, similar to the result of Kudo et al. (2022). There were no large differences in the value of the river discharge between the three deep learning models. Among them, SegFormer has the highest accuracy in predicting the water body. We concluded that the performance of these models is sufficient enough to estimate the river discharge.
On the other hand, some tricky images, such as low angle of incidence of sunlight (images taken at dawn or early evening), riverbeds strewn with small stones, or images with large shadows, are still difficult to be predicted accurately. We plan to add these tricky images to our dataset and perform some image preprocessing before the prediction in the future.