Using a deep learning model to quantify trash accumulation
Published in Computers, Environment and Urban Systems.
With growing understanding of trash impacts on aquatic habitats throughout the world, cities increasingly face regulatory requirements to reduce trash inputs to local waterways and the ocean, but they often rely upon insufficient monitoring data to prioritize and measure trash reduction effectiveness. We present an approach designed to make urban trash monitoring more cost-efficient and align the data collected with critical information needs of cities. We quantified urban trash accumulation along roadsides using vehicle mounted cameras and a deep convolutional neural network model to identify trash in the imagery captured. We compared the trash detection performance of three different models, with the best performing model (Mask R-CNN) achieving 91% recall, 83% precision, and 77% accuracy using data collected along 84 road segments in two California Cities. Trash detection model outputs were interpreted via a statistical model to relate the proportion of image pixels identified as trash to measured trash volumes. The resulting model estimates explained 67% of the variance in measured trash volumes collected on roadsides, which is more than double the variance explained by walking visual assessments. With vastly more efficient data collection compared to the visual assessments, deep learning based monitoring approaches can provide a stronger basis for understanding urban trash sources, changes over time, and cost-effective compliance with stormwater regulatory requirements.