Deep learning-based total suspended solids concentration classification of stream water surface images captured by mobile phone

• Published in: Environmental monitoring and assessment 2023-12, Vol.195 (12), p.1498-1498, Article 1498
• Main Authors: Hacıefendioğlu, Kemal, Baki, Osman Tuğrul, Başağa, Hasan Basri, Mete, Betül
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Atmospheric Protection/Air Quality Control/Air Pollution; Cell phones; Cellular telephones; Classification; Deep Learning; Earth and Environmental Science; Ecology; Ecotoxicology; Environment; Environmental Management; Environmental Monitoring - methods; Hydrologic processes; Image classification; Machine learning; Monitoring/Environmental Analysis; Neural networks; Neural Networks, Computer; Pollution control; Rivers; Sampling; Solid suspensions; Stream pollution; Stream water; Streams; Suspended particulate matter; Total suspended solids; Water; Watersheds
• ISSN: 0167-6369; 1573-2959
• DOI: 10.1007/s10661-023-12110-y

The continuous monitoring of total suspended solids (TSS) in streams plays an important role in the management of hydrological processes, and TSS is also a decisive parameter in the control of pollution in streams. Determination of TSS involves both a costly and time-consuming process. For this reason, there is a need to determine TSS more quickly, easily reliably, and economically. In this study, it is aimed to classify TSS concentration by using stream water surface images recorded with a mobile phone. For this purpose, the sampling studies carried out in the Lower Değirmendere Stream Watershed (Trabzon, northeast of Turkey) between June 2020 and May 2021, and the images of the stream water surface were taken and recorded with a mobile phone simultaneously with the sampling studies. The stream water surface images and TSS concentration were divided into four classification group, namely clear, turbid, very turbid, and extremely turbid and processed with deep learning method. More than 5500 stream water surface images and TSS concentration data were divided into groups as 70% training and the rest of test. This study uses eight convolutional neural network (CNN) models, including a custom CNN model designed for image classification, along with pre-existing architectures such as VGG-16, VGG-19, Inception-V3, Xception, DenseNet-121, MobileNet, and NASNet-mobile. The results revealed that the most effective ensemble model formation strategy was stacking with an accuracy rate of 87%, followed by weighted averaging, averaging, and majority voting. As a result, with the image of the stream water surface via the mobile phone camera, it can be easily determined which class the current TSS concentration belongs to instantly.