University of Twente Student Theses
Deep Learning-Based Semantic Segmentation for Detecting Marine Oil Spill
Ibrahim, Andi (2024) Deep Learning-Based Semantic Segmentation for Detecting Marine Oil Spill.
|
PDF
3MB |
| Abstract: | The rapid expansion of global maritime industries has increased the risk of oil spills, which can cause extensive marine pollution. Due to their rapid spread, oil spills often negatively impact marine biodiversity and coastal communities. Such accidents require rapid detection to mitigate their rapidly spreading negative impact. In recent years, deep learning (DL) techniques for oil spill detection (OSD) have advanced rapidly, particularly through the use of SAR images. However, obtaining a large number of labelled images that correspond to ground truth data of oil spill events is challenging. This limitation in input data presents a main obstacle to achieving optimal semantic segmentation results in the development of DL-based. This study investigates optimizing DL-based models by configuring relevant hyperparameters to accommodate limited datasets and adapt the oil spill appearance in SAR images for semantic segmentation in OSD. This study implements a Fully Convolutional Network-Dilated Kernel with 6 layers (FCN-DK6) model with fewer network parameters by optimizing the image preprocessing hyperparameter, specifically the patch size. Additionally, this study develops a hybrid CNN-XGBoost model based on convolutional networks from the FCN-DK6 model and a pre-existing XGBoost algorithm as an alternative solution in optimizing the FCN-DK6 model for OSD. This hybrid model has the potential to augment small datasets by generating multiple synthetic feature maps through its CNN component, which are then classified by XGBoost algorithm. The results show that the FCN-DK6 model despite having fewer network parameters can perform semantic segmentation in OSD on a limited dataset by optimizing a patch size to 512x512. This configuration achieves an optimal IoU accuracy of 51.03% for oil spills, outperforming smaller patch size settings. Furthermore, reducing the number of classes during training the model significantly improves the FCN-DK6 mode’s ability to segment oil spills by 98.80%. Meanwhile, the hybrid CNN-XGBoost model optimally performs OSD with semantic segmentation by configuring CNN filters and obtaining the optimal number of feature maps for feeding the XGBoost classifier. Combining optimal feature map settings with XGBoost hyperparameter tuning proves beneficial by achieving optimal performance with an IoU accuracy of 51.40% for OSD in a reduced class number scenario. While optimizing the FCN-DK6 model and developing the hybrid CNN-XGBoost model result in positive findings, the accuracy of OSD still needs more improvement to achieve reliable performance. This study provides scientific advancements by demonstrating the use of optimizing hyperparameters, particularly patch size in the DL-based model, to enhance OSD using a limited dataset. Additionally, the hybrid CNN-XGBoost model addresses input data limitations in OSD by generating multiple synthetic feature maps from a single input image to expand the dataset and fine-tune hyperparameters. These contributions advance deep learning techniques for marine environmental monitoring, particularly in data resource-limited scenarios. |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Subject: | 38 earth sciences, 43 environmental science |
| Programme: | Geoinformation Science and Earth Observation MSc (75014) |
| Link to this item: | https://purl.utwente.nl/essays/102148 |
| 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