Abstract Synthetic Aperture Radar (SAR) data are currently the most reliable resource for flood monitoring, though still subject to various uncertainties, which can be objectively represented with probabilistic flood maps. Moreover, the growing number of SAR satellites has increased the likelihood of observing a flood event from space through at least a single SAR image, but generalized methods for flood classification independent of sensor characteristics need to be developed, to fully utilize these images for disaster management. Consequently, a neuro-fuzzy flood mapping technique is proposed for texture-enhanced single SAR images. Accordingly, any SAR image is first processed to generate second-order statistical textures, which are subsequently optimized using a dimensionality reduction technique. The flood and non-flood classes are then modelled within a fuzzy inference system using Gaussian curves. Parameterization is achieved by training a neural network on the image through user-defined polygons. The results of the optimized texture-based neuro-fuzzy classification were compared against the performance of the SAR image alone and that of SAR enhanced with randomly selected texture features. This approach was tested for a COSMO-SkyMed SAR image at two validation sites, for which high resolution aerial photographs were available. An overall accuracy assessment using reliability diagrams demonstrated a reduction of 54.2% in the Weighted Root Mean Squared Error (WRMSE) values compared to the stand-alone use of SAR. WRMSE values estimated for the proposed method varied from 0.027 to 0.196. A fuzzy validation exercise was also proposed to account for the uncertainty in manual flood identification from aerial photography, resulting in fuzzy spatial similarity values ranging from 0.67 to 0.92, with higher values representing better performance. Results suggest that the proposed approach has demonstrated potential to improve operational SAR-based flood mapping.

中文翻译：

---

使用基于神经模糊纹理的方法实现基于 SAR 的洪水映射

摘要 合成孔径雷达（SAR）数据是目前洪水监测中最可靠的资源，但仍存在各种不确定性，可以用概率洪水图客观表示。此外，越来越多的 SAR 卫星增加了通过至少一张 SAR 图像从太空观测洪水事件的可能性，但需要开发独立于传感器特征的通用洪水分类方法，以充分利用这些图像进行灾害管理. 因此，针对纹理增强的单个 SAR 图像提出了一种神经模糊洪水映射技术。因此，首先处理任何 SAR 图像以生成二阶统计纹理，然后使用降维技术对其进行优化。然后使用高斯曲线在模糊推理系统中对洪水和非洪水类进行建模。参数化是通过用户定义的多边形在图像上训练神经网络来实现的。将优化的基于纹理的神经模糊分类的结果与单独的 SAR 图像的性能和使用随机选择的纹理特征增强的 SAR 的性能进行比较。该方法在两个验证站点针对 COSMO-SkyMed SAR 图像进行了测试，这些站点提供了高分辨率航拍照片。使用可靠性图进行的总体精度评估表明，与单独使用 SAR 相比，加权均方根误差 (WRMSE) 值降低了 54.2%。为建议的方法估计的 WRMSE 值从 0.027 到 0.196 不等。还提出了模糊验证练习，以解决航空摄影手动洪水识别的不确定性，导致模糊空间相似度值范围为 0.67 至 0.92，值越高表示性能越好。结果表明，所提出的方法已显示出改进基于操作性 SAR 的洪水测绘的潜力。