Machine Learning Based Water Analysis Using an Underwater Robot

Water Contamination is an important issue in many cases, for example aquaculture and aquaponics and dealing with it requires to recognize and identify it. While some approaches for water analysis exist, they either are non-spatial or only use 2D data. To solve that issue, this thesis deals with developing an approach that can transform spatial data into features and then use them with standard classifiers to classify different kinds of contamination, as well as introduce requirements for a robotic system, which is capable of gathering the data. o validate this approach, data was simulated in the form of multiple distributions exhibiting different properties to examine how well the approach handles them. Besides that, different sets of measurements were used with different spatial resolutions alongside a whole set of classifiers, and additional factors such as incomplete data and measurements deviating from their planned position were included. The classifier “Nearest Neighbors” proves to be the most potent for classification and delivers good results even for low-resolution data. Incompleteness and Deviation show effects with rising degree, and should be avoided, but the approach still delivers robust results, proving that with a robot capable of making and delivering precise measurements, the approach is viable.