VENµS raw images for vessel detection

The dataset utilized in this study comprises images capturing maritime vessels. It is a meticulously curated collection designed explicitly for training and assessing models focused on vessel detection and identification. The images within the dataset are sourced from satellite sensors, specifically the VENμS satellite imagery. VENμS offers multispectral data spanning from visible to near-infrared bands, boasting high spatial resolution (5.3 meters at Nadir) for detailed vessel analysis. The dataset encompasses images taken on 12 spectral bands, facilitating the examination of the vessel detection model's performance under diverse spectral conditions. These images depict the port of Ashdod, one of Israel's primary cargo ports, spanning from the beginning of 2018 to the conclusion of October 2020, totalling 284 multispectral images. For each multispectral image, three types of data are provided: raw data (non-coregistered multispectral image), coregistered images, and coregistered images with a mask (pixels set to zero) on the harbour to eliminate anchored vessels from the image. Coregistration involves a shift (X and Y) of each band to the fifth band (in the red). The majority of vessels in the dataset are annotated with boundary boxes (upper-left position, height, and width) along with AIS data, including vessel type, length, width, and status. These annotations serve as ground truth labels for training and evaluating models focused on vessel detection and classification. Furthermore, the dataset incorporates samples showcasing variations in environmental conditions, such as diverse sea states, lighting conditions, wake complexities, and the presence of clouds. This variability facilitates a comprehensive evaluation of the model's generalization capability and robustness across a spectrum of scenarios.