2022-2024 | Development of a Methodology for Automated Space Monitoring of Oil Spills Using Neural Network Technologies
- ИРН AP14872458
- Deadlines: 2022-2024 гг.
- PI: Nurseitov Daniyar, PhD, associate professor
- Project goal: Development of a Methodology for Automated Space Monitoring of Oil Spills Using Neural Network Technologies.
Project Tasks
2022: Create a database (DB) consisting of images with confirmed oil pollution and examples of non-target classes – the task of the year will be to review existing image analysis tools for detecting oil pollution; write guidelines for image selection and preprocessing parameters, and directly create a dataset of pollution and non-target classes.
2023: Develop and adapt an ML model for detecting oil pollution – to achieve this task, it is necessary to develop and compile the selected ML model and test it on the created dataset to determine and identify key and secondary features for identifying pollution or non-target classes.
2024: Develop a methodology for automated space monitoring of oil spills using neural network technologies – to achieve the goal, the preprocessing of images and additional parameters for pollution detection will be automated.
Expected results
2022: A database (dataset) of groups with reliable oil pollution and non-target classes, identified from satellite images of publicly available missions with annotated information for each event, has been formed.
2023: Compiled and tuned ML model for automatic detection of oil pollution.
2024: A methodology for automated space monitoring of oil spills based on neural network technologies with automated pre-processing of primary data has been developed.
Results Obtained
A database of reliable oil spills on land and non-target classes identified from Landsat satellite images for the period from 2006 to 2024 with annotated information for each event has been created. A dataset of marine oil spills and non-target classes was generated based on Sentinel-1 synthetic aperture radar (SAR) images for the period 2019-2024, including annotated images for each image. Based on the created datasets, convolutional neural network models were trained and configured for automatic detection of oil spills. Different models were compared on the training datasets. In the case of detecting oil spills on land, the model based on the Mask2former architecture showed the best results. For oil spills at sea, the DeepLabV3+ model was selected based on performance metrics. The developed methodology for automated space monitoring of oil spills on land includes automatic loading, pre-processing, processing of optical space images based on a deep learning model, post-processing to preserve the geospatial reference of data and posting the results on a geoportal. The methodology for automated space monitoring of oil spills on water also includes a set of sequential procedures for processing sea surface radar survey data, including the stages of automatic downloading, pre-processing and processing of fragments using neural network technologies, post-processing, expert verification of results and uploading results to a geoportal The developed automated monitoring technique based on neural network technologies demonstrates the effectiveness of using trained models to detect oil pollution and will allow analyzing large flows of information for the Caspian Sea and the vast territory of the Caspian region.
List of Publications
Sagatdinova G.N., NurseitovD.B. Processing of radar data from the Sentinel-1 satellite for identification of oil spills in the Caspian Sea in the GEE environment // Hydrometeorology and Ecology. – 2024. Vol. 1. – P. 100-109. – ISSN: 2789-6323. DOI:10.54668/2789-6323-2024-112-1-100-109 (in Russian).
Nurseitov D.B., Abdimanap G., Abdallah A., Sagatdinova G., Balakay L., Dedova T., Rametov N., Alimova A. ROSID: Remote Sensing Satellite Data for Oil Spill Detection on Land // Engineered Science. – 2024. – Т. 32. – С. 1348. http://dx.doi.org/10.30919/es1348 (Q1 процентиль Scopus – 88) (in English)