River state classification combining patch-based processing and CNN

• DOI: 10.1371/journal.pone.0243073
• Publication Date: 2020-12-03T19:11:20Z
• Authors: Takahiro Oga; Ryosuke Harakawa; Sayaka Minewaki; Yo Umeki; Yoko Matsuda; Masahiro Iwahashi
• Keywords: Connected-component labeling

This paper proposes a method for classifying the river state (a flood risk exists or not) from surveillance camera images by combining patch-based processing and convolutional neural network (CNN). Although CNN needs much training data, number of is limited because does not frequently occur. Also, include objects that are irrelevant to risk. Therefore, direct use may work well classification. To overcome this limitation, develops adjusting By increasing data via patch segmentation an image selecting patches relevant state, adjustment general CNNs classification becomes feasible. The proposed developed independently. yields practical merits any can be used according each user’s purposes, maintenance improvement component whole system easily performed. In experiment, defined as following problems using two datasets, verify effectiveness method. First, public dataset called Places classified with Muddy labels Clear labels. Second, in Nagaoka City, Japan captured when government announced heavy rain warning other images.