Mudflat aquaculture labeling for infrared remote sensing images via a scanning convolutional network

Mudflat areas, e.g. the enclosures of coastal inter-tidal regions, are sometimes used for breeding fish and other aquatic life, which is important for the aquaculture industry. As the difference of the wave reflectance between water and land structures of the infrared band is much higher than that of the visible band, infrared remote sensing technique is more suitable for automatically monitoring the mudflat aquaculture. This paper proposes a fast pixel-wise labeling method called scanning convolutional network (SCN) for mudflat aquaculture area detection with infrared remote sensing images. SCN improves the traditional fully convolutional network (FCN) by replacing convolution layers with scanning convolution modules (SCM) and a feature pyramid design, which simultaneously learns large scale sea-land environmental features and mudflat structure details with less computational costs. A set of Landsat-8 satellite images, with three visible bands and three infrared bands, are used to evaluate the proposed method. SCN shows a faster processing speed and a higher labeling accuracy than any other state of the art labeling methods.