Urban Flood Mapping with Residual Patch Similarity Learning

Urban Flood Mapping With Bitemporal Multispectral Imagery Via a Self-Supervised Learning Framework

“…Li et al [21] developed an active self-learning convolutional neural network (CNN) to classify the syntheticaperture-radar (SAR) image patches into three classes (i.e., non-flooded, flooded with buildings, and flooded without buildings). Peng et al [12], [13] designed a Siamese CNN model to evaluate the patch similarity for identification of flooded MS image patches. Song et al [29] and Sharma et al [30] proposed CNN-based models to map land cover with superior performance compared with pixel-based methods, especially in heterogeneous urban areas.…”

“…These aforementioned flood mapping studies, however, have focused on rural areas with relatively homogeneous image backgrounds. Meanwhile, flood extent mapping is insufficiently investigated in urban areas due to heterogeneous land cover and land use, low spatial resolution of MS imagery, and lack of flood extent ground truth datasets [12], [13].…”

“…Such inconsistency is a barrier to robust floodwater detection from multi-temporal RS images using data-driven machine learning models. Geometric and radiometric corrections are needed before surface reflectance imagery products can be useful for consistent spectral-based identification of flooded areas [12], [13].…”

“…Regarding the mapping strategy, both pixel-wise [4], [5], [16]- [20] and patch-wise [12], [13], [21] models have been proposed for flood mapping. Pixel-wise flood mapping methods assign each pixel of the input imagery a label (e.g., flooded or non-flooded) whereas patch-wise methods predict the class of an entire patch cropped from the imagery.…”

“…Both pixel-and patch-wise mapping have been explored for flood extent mapping, since each has unique advantages in specific scenarios. Patch-wise mapping has been widely used for flood mapping over heterogeneous urban areas [12], [13], [21], for several reasons. First, patch-wise mapping helps mitigate the impact of errors from multi-temporal image registration.…”