Patch-based CNN for real-time detection of still water in outdoor environments

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Water detection in open outdoor environments is an important requirement for many applications, such as aerial surveillance, public safety, environmental monitoring, and waste management. The existing water detection methods mostly focus on large water bodies. Detecting still water in small regions is a challenging task because water often appears similar to surrounding surfaces and has very limited texture. In this work, a lightweight patch-based convolutional neural network (CNN) model is presented for the detection of still water in real-time videos. Since no benchmark dataset is available for this problem, a small dataset consisting of videos recorded in different outdoor settings was created. Water and no-water patches of two different sizes were extracted from these videos and used to train a lightweight CNN that can detect small water regions in each frame. The trained CNN is then applied to real-time videos by dividing each frame into non-overlapping patches, classifying them as water or no-water, and combining the predictions to form the final water mask. The proposed method is evaluated on several videos and compared with classical segmentation methods, including active contour, color-based, graph cut, thresholding, Chan-Vese, K-Means, and SVM. The quantitative results show that the proposed model achieves an F-measure score between 95% and 96%, while operating at up to 24 frames/s on a standard CPU without GPU acceleration, performing better than other methods. A discussion on computational cost, convergence behaviour, and the limitations of the dataset is also included to support the feasibility of the proposed model. The results confirm that the proposed workflow is accurate, simple to implement, and suitable for detecting small water areas in different outdoor environments.