Machine Learning techniques applied to UAV imagery for macro-plastic detection in the fluvial environment

Abstract

Plastic ranks as the third most produced material globally by industry, following concrete and steel. A substantial quantity of the plastics manufactured around the world is now dispersed throughout the environment, particularly within aquatic ecosystems. Such dispersion can result in detrimental effects on flora, fauna, and human populations. Consequently, numerous efforts have been undertaken to develop methodologies and tools, either automatic or semi-automated, for detecting plastic pollution, aiming to facilitate its retrieval. This thesis emerges within this context, driven by the growing need to propose a methodology suitable for environmental monitoring capable of identifying macro-plastics (size > 5 mm) using UAV imagery and automatic classification methods, leveraging the well-recognized potential of machine learning tools in object detection applications.Considering the scarcity of available data on river environments and macro- plastics online, a significant step of this thesis involved programming and conducting data collection campaigns. Specifically, surveys were carried out using UAVs and mini-UAVs equipped with RGB, multi-spectral, and thermal cameras to obtain diverse and varied data-sets. The selected study areas are all located in Italy, pre- dominantly in the Tuscany region. These areas were chosen based on diverse characteristics such as vegetation, river width and depth, etc. Two different scenarios were defined: floating plastic on the river water surface; waste scattered along the riverbanks. The data-sets collected for the first case (floating plastic on the river water surface) comprised multi-spectral images (VIS, NIR, thermal), and a Machine Learning tool based on Random Forest classifiers was developed to accurately detect plastic objects. In this scenario, besides defining an effective workflow based on supervised pixel-based classification, it was particularly challenging to co-register multi- spectral images (VIS and NIR) with thermal images from diverse and asynchronous cameras. In fact, ideally, the sensors used for this type of acquisition should be synchronized with each other, but the available equipment (composed of the multi- spectral camera MAIA-S2 and the thermal camera DJI H20T) did not have these characteristics.In the pollution along the riverbanks-case (second case), the implemented strategy leverages high-resolution RGB images, a waste object detector based on recently developed Deep Learning methods (e.g. YOLO network), and a vision-based localization system. This combination substantially diminishes the positioning discrepancy of the UAV navigation system, thereby furnishing waste object position estimations accurate to the decimeter level, particularly for objects situated in proximity to identifiable locations on the map.The results obtained confirm the feasibility of the implemented approach and reflect the quality of the data collection and processing methods employed. A key finding of this research is that both methodological improvements and broadening of the scope of the study are necessary to establish a standardized procedure for macro-plastic identification and localization. Increasing the robustness and diversity of the datasets, as well as utilizing newer versions of Deep Learning algorithms, could be the initial steps towards this goal. Furthermore, among future developments, it is important to highlight the necessity for a comparative analysis of the advantages and disadvantages of the cases under consideration, along with potential integration of the implemented methodologies.