SANet: A Sea–Land Segmentation Network Via Adaptive Multiscale Feature Learning

Sea-land segmentation of remote sensing images is great significance to the dynamic monitoring coastlines. However, types objects in coastal zone are complex, and their spectra, textures, shapes, distribution features different. Therefore, sea-land for various coastlines still a challenging task. In this article, scale-adaptive semantic network, called SANet, proposed images. SANet has made two innovations on basis classic encoder-decoder structure. First, integrate spectral, textural, ground at different scales, we designed an adaptive multiscale feature learning module (AML) replace conventional serial convolution operation. The AML mainly contains extraction unit fusion unit. former can capture detailed information contextual from early stage, while latter adaptively fuse maps scales. Second, adopted squeeze-and-excitation bridge corresponding layers codec so that selectively emphasize weak boundaries. Experiments set Gaofen-1 demonstrated achieved more accurate results obtained sharper boundaries than other methods natural artificial