Deep learning (DL) methods have achieved promising results in high resolution (HR) remote sensing image (RSI) segmentation. While DL approaches rely heavily on large-scale training datasets, and a method trained on a dataset generally cannot be used to segment classes that the dataset does not contain. Existing RSI datasets mainly cover buildings, roads, vegetation, water, moving objects, etc. In this paper, we focus on a class of objects that have been rarely involved in previous datasets: offshore farms. The monitoring and management of offshore farms is important for the sustainable development of the aquaculture industry, and in this work, we provide an HR remote sensing dataset of offshore farms, named HROF. This dataset uses two types of sensors, i.e., multispectral and synthetic aperture radar (SAR). The multispectral data consists of 30 images with a resolution of about 1 m, including both raft and long line farms. The SAR data contains 41 images with a resolution of 3 m, in which the objects are breeding ponds. Objects in each image are manually annotated at the pixel-precise level. Furthermore, we provide baseline segmentation method and results on HROF and compare with mainstream methods. The dataset is available at https://github.com/FrontierQ/HROF.
|