[Paper Review] Waste Detection and Change Analysis based on Multispectral Satellite Imagery

Dávid Magyar, Máté Cserép+2 more|arXiv (Cornell University)|Mar 25, 2023

Municipal Solid Waste Management4 citations

TL;DR

This paper proposes a machine learning-based method using multispectral satellite imagery and Random Forest classification to detect illegal waste dumps (hot-spots) and floating waste blockages in rivers. The approach achieves accurate detection and change analysis on the Tisza River, with water-surface blockages showing the most reliable results, enabling scalable, cost-effective monitoring for environmental organizations.

ABSTRACT

One of the biggest environmental problems of our time is the increase in illegal landfills in forests, rivers, on river banks and other secluded places. In addition, waste in rivers causes damage not only locally, but also downstream, both in the water and washed ashore. Large islands of waste can also form at hydroelectric power stations and dams, and if they continue to flow, they can cause further damage to the natural environment along the river. Recent studies have also proved that rivers are the main source of plastic pollution in marine environments. Monitoring potential sources of danger is therefore highly important for effective waste collection for related organizations. In our research we analyze two possible forms of waste detection: identification of hot-spots (i.e. illegal waste dumps) and identification of water-surface river blockages. We used medium to high-resolution multispectral satellite imagery as our data source, especially focusing on the Tisza river as our study area. We found that using satellite imagery and machine learning are viable to locate and to monitor the change of the previously detected waste.

Motivation & Objective

• Address the growing environmental threat of illegal waste dumps in forests, rivers, and remote areas.
• Develop an automated, scalable solution for detecting plastic waste hot-spots and floating waste islands using remote sensing.
• Support waste collection organizations with repeatable, cost-effective monitoring tools using satellite data and machine learning.
• Enable change detection over time to track waste accumulation and movement in river systems.

Proposed method

• Utilized medium- and high-resolution multispectral satellite imagery from Sentinel-2 and PlanetScope missions.
• Applied Random Forest classification to detect waste based on spectral signatures across Blue, Green, Red, and Near-Infrared bands.
• Calculated spectral indices (e.g., NDVI, MNDWI) to enhance differentiation between waste and background surfaces like water and vegetation.
• Performed morphological transformations to refine detected waste regions and reduce noise.
• Implemented an automated alert system that compares daily satellite images to detect significant changes in waste coverage.
• Developed a web application prototype to visualize detected waste areas and track changes over time.

Experimental results

Research questions

• RQ1 Can multispectral satellite imagery combined with machine learning effectively detect illegal waste dumps in remote or inaccessible areas?
• RQ2 To what extent can spectral indices and machine learning distinguish floating plastic waste from water, vegetation, and shadows in river systems?
• RQ3 How effective is the method in identifying and monitoring changes in waste accumulation over time?
• RQ4 Can the system support operational waste collection by providing timely alerts for newly formed waste hot-spots or river blockages?

Key findings

• The method successfully detected waste hot-spots, including the Deponia Waste Management Centre in Hungary and Lake C˘alines,ti in Romania, with clear classification results.
• Water-surface river blockages were detected with high reliability, showing distinct red regions clearly separated from surrounding water and land.
• Execution time for processing high-resolution images (e.g., 6614×5981 pixels) reached up to 29 minutes and 41 seconds, with index computation being the main bottleneck.
• The system demonstrated feasibility for daily monitoring, with automated alerts triggered upon significant changes in detected waste areas.
• The Random Forest model performed best during spring and summer due to clearer image conditions, while winter and autumn images were often too cloudy for reliable use.
• The web application prototype successfully visualized waste extent and change trends, with plans for daily updates in the final version.