University of Twente Student Theses

Login

DumpingMapper : Illegal dumping detection from high spatial resolution satellite imagery

Sallander, J.G.D. (2023) DumpingMapper : Illegal dumping detection from high spatial resolution satellite imagery.

[img] PDF
1MB
Abstract:In the consumption-based society of today, illegally dumped waste constitutes a real issue. Illegal dumping comes with several environmental, health, and economic issues and thus should be reduced. Additionally, recycling dumped waste would be a great step towards a circular economy. Literature suggests that deep learning models, more specifically convolutional neural networks (CNNs), trained on high-resolution satellite imagery can be a possible solution to this problem. CNNs are successfully used to classify materials or objects on satellite imagery in similar fields. This project aims to find out how deep learning models can be used to classify illegally dumped materials on satellite imagery. Initially, a few state-of-the-art CNN models are trained to classify between just images that have dumping and images that do not. This is done at the hand of over 20.000 manually annotated satellite images of Cyprus, where ground truth data was collected. The Inception V3 CNN model was able to reach an overall classification accuracy of 83%, where the size of the dataset and training time had the largest impact on performance. Then, several CNN models are trained to classify two very distinct classes of waste material. This is done at the hand of a much smaller pre-existing dataset which is annotated with the required details of less than 600 images. This dataset was used as the training data while a subset of the earlier Cyprus dataset was re-annotated and used to evaluate the models. Using data augmentation techniques to slightly reduce the limitations of the small dataset, the Inception V3 CNN model was able to reach a classification accuracy of only 58%. While the approach used in this project does not produce confident performance figures for CNN models at classifying illegally dumped waste material, the earlier results and increased performance over different iterations of models suggest that with a different approach and implementation of new techniques, CNN deep learning models combined with high-spatial resolution satellite imagery might still pose a possible solution to illegal waste dumping in the future.
Item Type:Essay (Bachelor)
Clients:
CYENS - Centre of Excellence, Nicosia, Cyprus
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:38 earth sciences, 54 computer science
Programme:Creative Technology BSc (50447)
Link to this item:https://purl.utwente.nl/essays/94621
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page