A Deep Super-Resolution Mapping Network for Rice Paddy Extraction from Low-Resolution Remote Sensing Images

The segmentation of rice paddies using low-resolution images (LRIs) and the outputs to high-resolution maps (HRMs) is critical for continuous monitoring of crops. It is possible to conduct a ground survey, but it is both time-consuming and expensive. Deep image segmentation models like U-NETs offer suitable alternatives since these models can learn the "super-resolution" relationship between LRIs and HRMs. However, to learn this relationship, a large number of train LRI and HRM pairs must be obtained which may be prohibitively expensive. To reduce the number of training samples, new regularization terms derived from the level set functions [1] are added. As an example, the segmentation performance of a U-Net model was compared with and without the level set regularization terms (LSRTs) on the rice paddy extraction problem where Sentinel-2 images of size 32 × 32 pixels were used to generate rice paddy maps of size 256×256 pixels. When the regularization terms were employed, the Intersection-over-Union (IoU) and Rand Index (RI) of 0.719 and 0.716 were reported whereas the U-NET model without regularization terms can only obtain the IoU and RI of only 0.675 and 0.552, respectively. As a result, LSRTs enable the U-NET model to capture the fine borders at sub-pixel accuracy.