GEO TRACKING OF WASTE AND TRIGGERING ALERTS AND MAPPING AREAS WITH HIGH WASTE INDEX

Abstract

This article presents the use of automated 
machine learning for solving a practical 
problem of a real-life Smart Waste 
Management system. In particular, the 
focus of the article is on the problem of 
detection (i.e., binary classification) of an 
emptying of a recycling container using 
sensor measurements. Numerous datadriven methods for solving the problem 
were investigated in a realistic setting 
where most of the events were not actual 
empty ings. The investigated methods 
included the existing manually engineered 
model and its modification as well as 
conventional machines learning
algorithms. The use of machine learning 
allowed improving the classification 
accuracy and recall of the existing 
manually engineered model from 86:8 % 
and47:9 % to 99:1 % and 98:2 %
performing solution. This solution used a 
Random Forest classifier on a set of 
features based on the filling level at 
different given time spans. Finally, 
compared to the baseline existing
manually engineered model, the best 
performing solution also improved the 
quality of forecasts for emptying time of 
recycling containers.