Mapping Waste Piles in an Urban Environment Using Ground Surveys, Manual Digitization of Drone Imagery, and Object Based Image Classification Approach

Abstract There is wide recognition of the threats posed by open dumping of waste in the environment, however, tools to surveil interventions for reducing this practice are poorly developed. This study explores the use of drone imagery for environment surveillance. Drone images of waste piles were captured in a densely populated residential neighborhood in Malawi. Images were processed using the Structure for Motion Technique and partitioned into segments using Orfeo Toolbox. A total of 509 segments were manually labelled to generate data for training and testing a series of classification models. Four supervised classification algorithms (Random Forest, Artificial Neural Network, Naïve Bayes and Support Vector Machine) were trained, and their performances were assessed in terms of precision, recall and F-1 score. Ground surveys were also conducted to map waste piles using a GPS receiver and determine physical composition of materials on the waste pile surface. Differences were observed between the field survey done by transect walk and drone mapping. Drone mapping identified more waste piles than field surveys and for each waste pile, the spatial extent of waste piles was computed. Predictions from the binary random forest model were the highest performing (Precision: 0.98, Recall: 0.98, and F-score: 0.98). Drone mapping enabled identification of waste piles in areas that cannot be accessed during ground surveys, and further allows the quantification of total land surface area covered by waste piles. Drone imagery-based surveillance of waste piles thus has the potential to guide environmental waste policy and evaluate waste reduction interventions.