Abstract Continuous monitoring of water surfaces is essential for water resource management. This study presents a nonparametric unsupervised automatic algorithm for the identification of inland water pixels from multispectral satellite data using multidimensional clustering and a high-performance subsampling approach for large scenes. Clustering analysis is a technique that is used to identify similar samples in a multidimensional data space. The spectral information and derived indices were used to characterize each scene pixel individually. A machine learning approach with random subsampling and generalization through a Naive Bayes classifier was also proposed to make the application of complex algorithms to large scenes feasible. Accuracy was evaluated using an independent dataset that provides water bodies in 15 Sentinel-2 images over France acquired in different seasons and that covers a large range of water bodies and water colour types. The validation dataset covers a water surface of more than 1200 km2 (approximately 12 million pixels) including over 80,000 water bodies outlined using a semiautomatic active learning method, which were manually revised. The classification results were compared to the water pixel classification using three of the major Level 2A processors (MAJA, Sen2Cor and FMask) and two of the most common thresholding techniques: Otsu and Canny-edge. An input mask was used to remove coastal waters, clouds, shadows and snow pixels. Water pixels were identified automatically from the clustering process without the need for ancillary or pretrained data. Combinations using up to three water indices (Modified Normalized Difference Water Index-MNDWI, Normalized Difference Water Index-NDWI and Multiband Water Index-MBWI) and two reflectance bands (B8 and B12) were tested in the algorithm, and the best combination was NDWI-B12. Of all the methods, our method achieved the highest mean kappa score, 0.874, across all tested scenes, with a per-scene kappa ranging from 0.608 to 0.980, and the lowest mean standard deviation of 0.091. Standard Otsu's thresholding had the worst performance due to the lack of a bimodal histogram, and the Canny-edge variation achieved an overall kappa of 0.718 when used with the MNDWI. For water masks provided by generic processors, FMask outperformed MAJA and Sen2Cor and obtained an overall kappa of 0.764. In-depth analysis shows a quick drop in performance for all of the methods in identifying water bodies with a surface area below 0.5 ha, but the proposed approach outperformed the second best method by 34% in this size class.

中文翻译：

---

Sentinel-2 图像多维层次聚类的自动水检测以及与 Level 2A 处理器的比较

摘要 水面的连续监测对于水资源管理至关重要。本研究提出了一种非参数无监督自动算法，用于使用多维聚类和大场景的高性能子采样方法从多光谱卫星数据中识别内陆水域像素。聚类分析是一种用于在多维数据空间中识别相似样本的技术。光谱信息和派生指数用于单独表征每个场景像素。还提出了一种通过朴素贝叶斯分类器进行随机子采样和泛化的机器学习方法，以使复杂算法在大型场景中的应用变得可行。使用独立数据集评估准确性，该数据集提供了在法国不同季节获得的 15 张 Sentinel-2 图像中的水体，涵盖了大量水体和水色类型。验证数据集覆盖了超过 1200 平方公里（约 1200 万像素）的水面，包括使用半自动主动学习方法勾勒出的 80,000 多个水体，这些水体经过手动修正。使用三个主要的 2A 级处理器（MAJA、Sen2Cor 和 FMask）和两种最常见的阈值技术：Otsu 和 Canny-edge，将分类结果与水像素分类进行比较。输入掩码用于去除沿海水域、云层、阴影和雪像素。水像素是从聚类过程中自动识别的，无需辅助或预训练数据。在算法中测试了使用最多三个水指数（修正归一化差值水指数-MNDWI、归一化差值水指数-NDWI 和多波段水指数-MBWI）和两个反射带（B8 和 B12）的组合，最佳组合是 NDWI -B12。在所有方法中，我们的方法在所有测试场景中实现了最高的平均 kappa 分数，0.874，每个场景的 kappa 范围从 0.608 到 0.980，最低的平均标准偏差为 0.091。由于缺乏双峰直方图，标准 Otsu 阈值的性能最差，当与 MNDWI 一起使用时，Canny-edge 变化实现了 0.718 的整体 kappa。对于通用处理器提供的水面罩，FMask 的表现优于 MAJA 和 Sen2Cor，并获得了 0.764 的整体 kappa。