Integrating neural network approaches with remote sensing for detection and prediction of oil contamination

Effective monitoring and remediation of hydrocarbon contamination are crucial for environmental protection. This study explores applying a neural network model with Sentinel-2 satellite imagery to detect and classify oil contamination in Kuwait's Greater Burgan Oil Field. We analyzed over 200 Sentinel-2A images covering approximately 195.43 km². We incorporated ground truth data from 180 soil samples across six systematically sampled blocks. The dataset was processed into 4320 image patches (2808 for training, 648 for validation, 864 for testing). To minimize spatial autocorrelation and ensure robust validation, we employed a block-based partitioning approach, spatially separating training and testing areas by at least 1 km. The model achieved promising results in identifying contaminated areas with 97.054 % overall accuracy and 96.156 % precision. It effectively distinguishes between wet oil lakes, dry oil lakes, and oil-contaminated piles, with particularly strong performance in detecting the latter (99.92 % recall). The model's ability to delineate wet and dry oil lakes is more limited (74.362 % recall). Quantitative evaluations include F1 scores (up to 91.47 %) and IoU (up to 84.28 %). Incorporating spectral indices enhances detection of subtle variations in spectral signatures associated with oil contamination. Remediation priority maps derived from model outputs offer a basis for prioritizing intervention areas. While integrating neural networks with remote sensing techniques does show promise, further validation across diverse geographical and contamination scenarios is necessary to enhance the model's generalizability.