Throughout the world, the likelihood of floods and managing the associated risk are a concern to many catchment managers and the population residing in those catchments. Catchment modelling is a popular approach to predicting the design flood quantiles of a catchment with complex spatial characteristics and limited monitoring data to obtain the necessary information for preparing the flood risk management plan. As an important indicator of urbanisation, land use land cover (LULC) plays a critical role in catchment parameterisation and modelling the rainfall–runoff process. Digitising LULC from remote sensing imagery of urban catchment is becoming increasingly difficult and time-consuming as the variability and diversity of land uses occur during urban development. In recent years, deep learning neural networks (DNNs) have achieved remarkable image classification and segmentation outcomes with the powerful capacity to process complex workflow and features, learn sophisticated relationships and produce superior results. This paper describes end-to-end data assimilation and processing path using U-net and DeepLabV3+, also proposes a novel approach integrated with the clustering algorithm MeanShift. These methods were developed to generate pixel-based LULC semantic segmentation from high-resolution satellite imagery of the Alexandria Canal catchment, Sydney, Australia, and assess the applicability of their outputs as inputs to different catchment modelling systems. A significant innovation is using the MeanShift clustering algorithm to reduce the spatial noise in the raw image and propagate it to the deep learning network to improve prediction. All three methods achieved excellent classification performance, where the MeanShift+U-net has the highest accuracy and consistency on the test imagery. The final suitability assessment illustrates that all three methods are more suitable for the parameterisation of semi-distributed modelling systems rather than the fully distributed modelling systems, where the MeanShift+U-net should be adopted for image-based impervious area extraction of urban catchment due to its superior prediction accuracy of 98.47%.

在全世界，洪水的可能性和相关风险的管理是许多集水区管理者和居住在这些集水区的人口的关注点。流域建模是一种流行的方法，用于预测具有复杂空间特征和有限监测数据的流域的设计洪水分位数，以获得制定洪水风险管理计划所需的信息。作为城市化的重要指标，土地利用土地覆被（LULC）在流域参数化和降雨-径流过程建模中发挥着关键作用。随着城市发展过程中土地利用的可变性和多样性，从城市集水区的遥感图像中数字化 LULC 变得越来越困难和耗时。最近几年，深度学习神经网络 (DNN) 已经取得了显着的图像分类和分割结果，具有处理复杂工作流程和特征、学习复杂关系并产生卓越结果的强大能力。本文描述了使用 U-net 和 DeepLabV3+ 的端到端数据同化和处理路径，还提出了一种与聚类算法 MeanShift 相结合的新方法。开发这些方法是为了从澳大利亚悉尼亚历山大运河流域的高分辨率卫星图像生成基于像素的 LULC 语义分割，并评估其输出作为不同流域建模系统输入的适用性。一项重要的创新是使用 MeanShift 聚类算法来减少原始图像中的空间噪声并将其传播到深度学习网络以改进预测。这三种方法都取得了出色的分类性能，其中 MeanShift+U-net 在测试图像上具有最高的准确性和一致性。最终的适用性评估表明，这三种方法都更适合半分布式建模系统的参数化，而不是全分布式建模系统，其中应采用 MeanShift+U-net 进行基于图像的城市流域不透水面积提取，因为98.47%的预测准确率。其中 MeanShift+U-net 在测试图像上具有最高的准确性和一致性。最终的适用性评估表明，这三种方法都更适合半分布式建模系统的参数化，而不是全分布式建模系统，其中应采用 MeanShift+U-net 进行基于图像的城市流域不透水面积提取，因为98.47%的预测准确率。其中 MeanShift+U-net 在测试图像上具有最高的准确性和一致性。最终的适用性评估表明，这三种方法都更适合半分布式建模系统的参数化，而不是全分布式建模系统，其中应采用 MeanShift+U-net 进行基于图像的城市流域不透水面积提取，因为98.47%的预测准确率。