*3.4. JRC Global Surface Water Products*

This study focused on depressional wetlands that, by definition, are not permanent water, and often change inundation status quickly due to climate variability. We used the JRC product to differentiate wetlands from permanent water bodies across the entire study area. The Joint Research Centre's Global Surface Water (JRC GSW) product contains the surface water's spatial and temporal distribution at 30 m resolution. The product provides different characteristics of surface water, including occurrence, intensity, seasonality, recurrence, transitions, and maximum water extent [36]. The JRC GSW data were generated using more than 3 million scenes from various Landsat missions (Landsat 5, 7, and 8) between 1984 to 2019. The pixels were classified into water and non-water classes using an expert system. JRC GSW presents results each month for the entire period (1984–2019) for change detection. We defined permanent water bodies as those classified as water in >90% of the observations within the period (1984–2019), and filtered those pixels from the study. The permanent wet pixels were excluded from the final results to map the surface waters that only belong to wetlands.

**Figure 4.** An RGB image of Sentinel-2 over a portion of the study area (**A**). NDWI time series in two small potholes in the study area (**B**), and showing significant temporal variations in surface water (**C**).

#### **4. Methods**

We developed an open-source process in GEE based on machine learning algorithms and multisensory remote sensing data for wetlands identification, as follows. First, a total of 895 ground truth points for 2016, including inundated wetlands and non-wetland classes, were randomly divided into two subsets of training (comprising 637 data points) and testing (comprising 258 data points). The training subset was used for training the machine learning algorithms, and the testing subset was withheld from the model, and used for the accuracy assessment. We created a multisensory band composite by integrating Sentinel-1 SAR data to selected Sentinel-2 high-resolution bands (Figure 3; Table 1). We used this Sentinel-1 and Sentinel-2 composite as predictors in the classification. We evaluated two machine learning algorithms, random forest (RF) and support vector machine (SVM), to establish a relationship between the multisensory composite bands as predictors and the training ground truth data. The optimum model (the model with the highest accuracy for classifying testing data) was used to classify the multisensory composite into two classes of wetlands and non-wetland pixels to identify wetlands in our study area. The generalizability of the optimum model was tested again using an additional 2231 ground truth points from a novel year, 2017.

Additionally, we tested the method by performing an accuracy assessment on small vegetated and small non-vegetated wetlands (see explanation below). Next, we excluded the permanent water bodies from the map using the JRC products as described above. Finally, we mapped the emergent vegetation within the identified wetlands using Sentinel-2-derived NDVI. We describe the details of the adopted methodology below. Figure 5 shows the workflow of the method.

**Figure 5.** Flowchart showing the main steps that were used in this study for mapping wetlands surface water.
