University of Twente Student Theses
Mapping road pavement quality from optical satellite imagery using machine learning
Gebreegziabher, Bisrat Araya (2021) Mapping road pavement quality from optical satellite imagery using machine learning.
|
PDF
4MB |
| Abstract: | Food loss occurring along the supply chain poses a major challenge in sustaining global food security. While agricultural production has improved significantly over the recent years, the facilities to manage this production have not kept up. This insufficiency results in post-harvest losses that occur after the harvesting of agricultural products. Post-harvest losses are prevalent issues in developing countries, thwarting the efficiency of agricultural food supply chains. Transportation has a substantial role in these losses since it is a vital link in the post-harvest chain. Particularly in developing regions, where road transport is the typical linkage, there is a decisive necessity to ensure the quality of transport facilitation. Ensuring quality in this sense means that the condition of roads has to be monitored, maintained, and rehabilitated. However, due to the lack of sufficient resources, these activities are not undertaken regularly. This aspect has resulted in the prevalence of poor-quality road that induces in-transit damages to perishable agricultural products such as tomatoes. This study argues that spatial road quality information is a valuable tool in addressing these challenges. More importantly, enabling the convenient accessibility of this information is vital for resource strained regions such as Sub-Saharan Africa. Towards this goal, this research investigated the potential of mapping road pavement quality from freely accessible optical satellite imagery using machine learning methods. Accordingly, shallow and deep learning models were developed to extract road quality information from Sentinel-2 satellite imagery using reference data collected for a corridor running from Accra (Ghana) to Ouagadougou (Burkina Faso) with crowdsensing technology. The results were encouraging in realizing the use of such a data source for convenient access to road pavement quality information. The deep learning model, i.e., U-Net, reported an F1-score of 37.93% and an IoU of 32.40%, outperforming the shallow ML alternative in the form of random forest. The inherent data imbalance prevents comparison with conventional segmentation task performance. The results, however, were comparable to analogous road extraction projects that utilized Sentinel-2 images. The study also contrasted the use of Sentinel-2 imagery to that of Planet imagery data to assess the relative potential of Sentinel-2 imagery in the task. The results showed that Sentinel-2 images were more suitable than the Planet ones in the pixel-wise classification of road pavement quality (RPQ). Furthermore, a three-class RPQ classification model was presented to resolve the ambiguity surrounding severity classes. With an F1-score of 53.65% and an IoU of 46.03%, this model performed substantially better. Alternative to this approach, a flexible modeling paradigm based on probabilistic threshold moving was also explored. Aided with heuristics of precision-recall tradeoff and the probabilistic nature of ML model predictions, the study showed that predictions of the models could be molded to suit the utility desired. |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Programme: | Spatial Engineering MSc (60962) |
| Link to this item: | https://purl.utwente.nl/essays/89012 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page