Instance Segmentation of Water Body from Aerial Image using Mask Region-based Convolutional Neural Network

Sangdaow Noppitak, Sarayut Gonwirat, and Olarik Surinta

Abstract

Land use is constantly changing, and water plays a critical role in the process. If changes are noticed quickly or are predictable, land use planning and policies can be devised to mitigate almost any problem. Accordingly, researchers present a mask region-based convolutional neural network (Mask R-CNN) for water body segmentation from aerial images. The system’s Aerial image water resources dataset (AIWR) was tested. The AIWR areas were agricultural and lowland areas that require rainwater for farming. Many wells were spotted throughout the agricultural areas. The AIWR dataset presents two types of data: natural water bodies and artificial water bodies. The two different areas appear as aerial area images that are different in color, shape, size, and similarity. A pre-trained model of Mask R-CNN was used to reduce network learning time. ResNet-101 was used as backbone architecture. The information gathered in the learning process is limited, and only 720 pictures were produced, Researchers used data augmentation to increase the amount of information for training by using affine image transformation, including scale, translation, rotation, and shear. The experiment found that mask R-CNN architecture can specify the position of the water surface. Measuring method in this case is mAP value. The mAP value is at 0.30 without data augmentation. However, if using the R-CNN mask with data augmentation, the mAP value increased to 0.59.

Keywords — Instance Segmentation; Water Body; Aerial Image; Mask R-CNN; Transfer Learning.

Read article — https://dl.acm.org/doi/abs/10.1145/3388176.3388184

Conclusion

In this paper evaluated the accuracy of instance segmentation by Mask R-CNN together with data augmentation. The mAP values were used as the measuring method. This research tested with aerial images of water resources dataset (AIWR). The areas are the lowlands which require rainwater for farming. The challenges of AIWR dataset the collection of 2 types of water bodies: natural water bodies and artificial water bodies. The two types of data are different in color, shape, size, and similarity. This paper used a pre-trained model to reduce learning time of the Mask R-CNN. This research has shown that the mask R-CNN architecture combined with data augmentation can identify the water surface using the mAP value for measurement. The value was up to 0.59. It is almost two times greater than not using data augmentation method.

In future work, because the data tested is aerial photography obtained from Bing map, only RGB colors can be evaluated. If other research can use data from satellites, such as Landsat, which has a band specifically for water analysis, the result of an analysis of water bodies with different color might give higher accuracy. Any new architecture suitable for water body analysis might be used to expect an even higher accuracy rate.