Highly destructive flood disasters have occurred frequently recently. Related to this, accurate mapping of flood areas is a necessary undertaking that helps to understand the temporal and spatial evolution patterns of floods. Thus, this paper proposes a novel, unsupervised multi-scale machine learning (ML) approach for urban flood mapping with SAR images from the perspective of information mining and fusion. Considering the complexity of surface objects in urban scenes, the proposed approach first extracts and fuses multiple types of features, such as polarization, pseudo-color, and spatial features, from pre-flood and post-flood SAR images to enhance distinguishability of water bodies. In particular, some new pseudo-color features are constructed here for SAR images through pseudo-color synthesis and color space transformation. On this basis, a flood probability map (FPM) is generated, and multi-scale superpixel segmentation is performed on it. Then, an ML-based unsupervised classification model assisted by uncertainty analysis based on the Gaussian mixture model is designed and implemented for flood mapping at different segmentation scales. Finally, guided by the minimum uncertainty, an adaptive fusion strategy of multi-scale information is proposed to integrate the flood mapping results at different scales for producing the final flood map. The proposed approach is unsupervised, and can minimize the mapping uncertainty to improve mapping accuracy and reliability. These characteristics of the proposed approach make it practical. The results of comparative experiments demonstrate that the proposed approach is effective and has certain advantages over existing methods, especially in reducing false detections and correctly identifying the categories of uncertain pixels in flood mapping. Furthermore, the experimental results also indicate that the pseudo-color features constructed here also help enhance flood mapping accuracy.

中文翻译：

---

通过充分挖掘和自适应融合 Sentinel-1 图像的极化和空间信息进行城市洪水测绘


最近，极具破坏性的洪灾灾害频发。与此相关，准确绘制洪水区域地图是一项必要的工作，有助于了解洪水的时间和空间演变模式。因此，本文从信息挖掘和融合的角度提出了一种新颖的、无监督的多尺度机器学习 （ML） 方法，用于使用 SAR 图像进行城市洪水制图。考虑到城市场景中地表对象的复杂性，所提方法首先从洪水前和洪水后的 SAR 图像中提取并融合多种类型的特征，例如极化、伪彩色和空间特征，以增强水体的可区分性。特别是，这里通过伪彩色合成和颜色空间变换为 SAR 图像构建了一些新的伪彩色特征。在此基础上，生成洪水概率图 （FPM），并对其进行多尺度超像素分割。然后，设计并实现了基于高斯混合模型的基于ML的无监督分类模型，并辅以不确定性分析，用于不同分割尺度的洪水映射。最后，在最小不确定性的指导下，提出了一种多尺度信息的自适应融合策略，将不同尺度的洪水制图结果进行整合，以生成最终的洪水图。所提出的方法是无人监督的，可以最大限度地减少映射的不确定性，从而提高映射的准确性和可靠性。所提出的方法的这些特点使其具有实用性。 对比实验结果表明，所提方法有效，与现有方法相比具有一定的优势，特别是在减少误检和正确识别洪水映射中不确定像素的类别方面。此外，实验结果还表明，这里构建的伪彩色特征也有助于提高洪水制图的准确性。