Fast and Automatic Data-Driven Thresholding for Inundation Mapping with Sentinel-2 Data

Abstract

Satellite data offer the opportunity for monitoring the temporal flooding dynamics of seasonal wetlands, a parameter that is essential for the ecosystem services these areas provide. This study introduces an unsupervised approach to estimate the extent of flooded areas in a satellite image relying on the physics of light interaction with water, vegetation and their combination. The approach detects automatically thresholds on the Short-Wave Infrared (SWIR) band and on a Modified-Normalized Difference Vegetation Index (MNDVI), derived from radiometrically-corrected Sentinel-2 data. Then, it combines them in a meaningful way based on a knowledge base coming out of an iterative trial and error process. Classes of interest concern water and non-water areas. The water class is comprised of the open-water and water-vegetation subclasses. In parallel, a supervised approach is implemented mainly for performance comparison reasons. The latter approach performs a random forest classification on a set of bands and indices extracted from Sentinel-2 data. The approaches are able to discriminate the water class in different types of wetlands (marshland, rice-paddies and temporary ponds) existing in the Doñana Biosphere Reserve study area, located in southwest Spain. Both unsupervised and supervised approaches are examined against validation data derived from Landsat satellite inundation time series maps, generated by the local administration and offered as an online service since 1983. Accuracy assessment metrics show that both approaches have similarly high classification performance (e.g., the combined kappa coefficient of the unsupervised and the supervised approach is 0.8827 and 0.9477, and the combined overall accuracy is 97.71% and 98.95, respectively). The unsupervised approach can be used by non-trained personnel with a potential for transferability to sites of, at least, similar characteristics.

Automatic Inundation Mapping Using Sentinel-2 Data Applicable to Both Camargue and Doñana Biosphere Reserves

“…Commonly used indices include the Normalized Difference Water Index (NDWI) [15,16,19], Modified NDWI [14,[18][19][20], and Automated Water Extraction Index [17][18][19]21]. Several approaches use information from Shortwave infrared (SWIR) spectral ranges to identify shallow inundated wetland areas, since it is less sensitive to sediment-filled waters and, hence, more efficient for registering the boundaries between water and dry areas in shallow wetlands [13,[22][23][24]. Automatic thresholding approaches can be applied to different areas and are computationally inexpensive, but they may wrongly classify dark objects (i.e., shadows and buildings) as water when their spectral characteristics are similar [25].…”

“…Automatic thresholding approaches can be applied to different areas and are computationally inexpensive, but they may wrongly classify dark objects (i.e., shadows and buildings) as water when their spectral characteristics are similar [25]. Automatic thresholding approaches are distinguished into: (a) global approaches [15,17,19,20,26], which estimate thresholds based on the histogram analysis of the complete image, and (b) local thresholding approaches [23], which estimate local thresholds for image subsets containing high percentages of pixels belonging to the water and non-water classes, and then may take into consideration subsets' thresholds to estimate an overall threshold. Local thresholding approaches overcome the incapability of global approaches to estimate an optimal histogram threshold when the class proportions within the image are imbalanced [27].…”

“…Most of the previously mentioned approaches aim to detect water in open-water bodies such as rivers, lakes, reservoirs, and watersheds [17,19,20,30,33,37,43], while a part of them focuses on wetland areas [13,[22][23][24].…”

“…This study focuses on wetland areas and examines alternative approaches of the original automated local thresholding approach presented in Reference [23] with the objective to suggest alternative approaches, which may produce credible results for both Camargue and Doñana wetlands, i.e., be identified as possibly applicable to further wetland areas and biomes. Each examined alternative approach relies on a specific band or band combination, acknowledged as effective by the underlying physics, and a specific approach for estimating splitting thresholds.…”

Wetland Classification Based on a New Efficient Generative Adversarial Network and Jilin-1 Satellite Image

“…Supervised methods, which are extensively used in wetland classification, have received improved performance by making full use of the prior information. A lot of supervised methods, such as support vector machine (SVM) [16,17] and random forest (RF) [18], have been reported in the literature. However, supervised methods always require many labeled samples to train satisfactory classifiers and a large number of unlabeled samples are ignored.…”