S1S2-Water: A Global Dataset for Semantic Segmentation of Water Bodies From Sentinel- 1 and Sentinel-2 Satellite Images

Abstract

This study introduces the S1S2-Water dataseta global reference dataset for training, validation and testing of convolutional neural networks for semantic segmentation of surface water bodies in publicly available Sentinel-1 and Sentinel-2 satellite images. The dataset consists of 65 triplets of Sentinel-1 and Sentinel-2 images with quality checked binary water mask. Samples are drawn globally on the basis of the Sentinel-2 tile-grid (100 x 100 km) under consideration of predominant landcover and availability of water bodies. Each sample is complemented with metadata and Digital Elevation Model (DEM) raster from the Copernicus DEM. On the basis of this dataset we carry out performance evaluation of convolutional neural network architectures to segment surface water bodies from Sentinel-1 and Sentinel-2 images. We specifically evaluate the influence of image bands, elevation features (slope) and data augmentation on the segmentation performance and identify best-performing baseline-models. The model for Sentinel-1 achieves an Intersection Over Union of 0.845, Precision of 0.932 and Recall of 0.896 on the test data. For Sentinel-2 the best model produces an Intersection Over Union of 0.965, Precision of 0.989 and Recall of 0.951 respectively. We also evaluate the performance impact when a model is trained on permanent water data and applied to independent test scenes of floods.

Flood Detection with SAR: A Review of Techniques and Datasets

“…They are routinely adopted for assessing classification capabilities of flooding detection algorithms, regardless of the methodology adopted. Indeed, the performance of probabilistic and fuzzy methods can be assessed using such metrics once binary classification maps are derived by assigning a probability threshold, typically set at 0.5, as it is performed, e.g., in [186]. This procedure is typically referred to as defuzzification in the context of fuzzy methods.…”

“…Consequently, strict time requirements may take precedence over accuracy. Computational complexity can be evaluated via the algorithm throughput measured as the average number of processed pixels per second [186].…”

“…It is worth mentioning that the most recent datasets for flood detection include both SAR and MS images. Indeed, it is well recognized that such sensors offer complementary features for detection of flooded areas, and recent works have proved that the best accuracy is achieved by a combined use of both kinds of data [121,186]. A proper scenario-based weighting of the two types of inputs is envisaged in [121].…”

“…The S1S2-Water is a global dataset conceived for the semantic classification of water bodies in S1 and S2 imagery [186]. It comprises over 100,000 non-overlapping patches, each sized 256 × 256, for each sensor.…”

“…Each patch is complemented by a corresponding DEM tile derived from the Copernicus DEM and quality-controlled binary water mask, as well as other metadata. The data samples cover the entire globe and are available for download at https://zenodo.org/records/8314175 (accessed on 21 January 2024) along with the code of the method proposed in [186].…”