CNN based Ocean Surface Feature Classification from Satellite Imagery

Abstract

Oil leaks are regarded as one of the most serious threats to aquatic and maritime ecosystems. Effective monitoring, ship tracking, and precise oil spill detection are critical for responding agencies to act appropriately while minimizing pollution damage and avoiding further destruction. Rapid detection of oil spills is critical for mitigating the effects and avoiding further disruption. Because of their broad range, satellites and tracking aircraft are used for marine supervision. Synthetic aperture radar (SAR) sensors are commonly employed for this application due to their ability to operate effectively under both temperature and lighting environments. While SAR sensors can easily detect dark marks correlated with oil spills, distinguishing them from look-alikes is a difficult task. A variety of ways have been suggested to locate and recognize these dark spots. Satellites used to collect data have resulted in the absorption of massive volumes of remote sensing data into networks, but analysing the data with human effort is a time-consuming and labour-intensive process. As a result, contemporary research proposes using machine learning as a replacement for conventional approaches in conceptual frameworks such as pattern segmentation, image recognition, and object detection. The majority of them make use of personalized datasets that are not always created equally in terms of efficiency. Furthermore, in most situations, a single mark is applied to the overall SAR image, which makes editing complicated circumstances or retrieving additional information from the portrayed material difficult. These shortcomings are resolved with the assistance of various deep neural networks. This research examines the success of various pix2pix GAN prototypes on the MKLAB oil spill datasets. When a tweaked U-net segmentation system is employed as a generator in pix2pix, the findings reveal a substantial result for SAR image segmentation as compared to those on the test set.