2025 – Initial dataset creation using a limited set of data through the collection, preprocessing, annotation, and creation of thematic layers with user-based or semi-automated classification of high-resolution satellite images containing oil-contaminated soil. Testing of the initial dataset on existing artificial neural network architectures.

2026 – Creation of a full-scale dataset ensuring data completeness and representativeness through the use of a large number of high-resolution images. Iterative improvement of the full-scale dataset based on testing on existing artificial neural network architectures.

2027 – Development of a new artificial neural network architecture designed for recognizing satellite multi-channel images with more than four spectral channels, taking into account the specific characteristics and features of each channel. Training, testing, and optimization of the artificial neural network model for recognizing oil-contaminated soil, including performance metric calculations and comparison with existing architectures.

2025 – Initial dataset will be created, containing archival high-resolution satellite images and annotated images with surface classes, including oil contamination. Testing of the initial high-resolution satellite image dataset on existing neural network architectures proven successful for similar tasks.

2026 – Initial dataset will be expanded with additional archival high-resolution satellite images capturing oil-contaminated soil cases. A full-scale dataset of annotated images with various types of surface classes, including oil-contaminated soil, will be created.

2027 – A new neural network architecture will be developed for recognizing satellite multi-channel images with more than four spectral channels. Training, testing, and optimization of the developed neural network model for oil contamination recognition will be carried out, along with comparison to existing architectures.

In 2025, an initial dataset was created, consisting of archived PlanetScope satellite imagery acquired over the period 2017–2024 with a spatial resolution of 3 m, along with annotated images containing land cover classes, including oil contamination. The main focus in constructing the dataset was the accurate identification of the “oil-contaminated area” class. To improve the accuracy of oil spill delineation during training, results from previous studies and the reference ROSID (Remote Oil Spill Identification Dataset), based on medium-resolution Landsat satellite data (30 m), were used. The initial dataset was tested on several artificial neural network architectures, such as DeepLabv3+ and Mask2Former. Model performance was evaluated using Accuracy, Precision, Recall, and F1-score metrics.

None.