Preliminary DEM Accuracy Assessment

Preliminary DEM Accuracy Assessment

A preliminary DEM accuracy assessment was designed to test how well the void-filled 90m SRTM DEM was at deriving flood depth and extent from ground water data. Gauge derived flood depth and extent calculations were obtained by subtracting terrain elevation from a gridded water level surface, retaining all grid cells with a depth greater than zero and setting all other grid cells to zero (Werner, 2001). This analysis captures where ground water, represented by a rubber like sheet and obtained via kriging as described above, breaks the surface which represents flood inundation.

Landsat imagery provide aerial snapshots of local conditions across the globe and are commonly used to classify surface water using Normalized Difference Water Index (NDWI) techniques. NDWI grids are created by:

because mid–infrared helps in segregating water from land since water absorbs while land reflects in this band. Four separate LS tiles (Supplementary Material Figure 1) are needed to capture and get a complete view of most of the country. Within each one of the four regions, multiple LS tiles were obtained to increase the sample size for the preliminary DEM accuracy assessment. A total of 15 Landsat 5 Thematic Mapper tiles were selected that met the criteria of having minimal cloud cover and matching temporal resolution with groundwater data within both wet and dry seasons. Four tiles were obtained in the south region, 5 tiles were obtained in the southwest region, 4 were tiles obtained in the Sylhet region, and 3 tiles were obtained in the west region. For each tile in each region, surface water was classified using NDWI methodology and then quantified in square kilometers (Supplementary Material Table 1). The final surface water area calculations from the Landsat 5 TM imagery were used as a control for testing the accuracy of the SRTM DEM in the preliminary accuracy assessment.

Groundwater gauge data that had matching temporal resolution with the Landsat 5 TM imagery were clipped by that LS tiles boundary and were then used to test the SRTM DEM’s accuracy. Flood depth was derived from gauge data using techniques described above, using the raw SRTM DEM terrain elevation, and then flood extent was quantified in square kilometers and recorded in Supplementary Material Table 2. Supplementary Material Table 2 compares gauge data derived and LS derived flood extent in km2.

Because SRTM DEM data have an absolute vertical accuracy ≤ 16m, we adjusted the DEM level to minimize the difference between surface water area derived from gauge data and Landsat imagery. Vertical accuracy of grid cell elevation is a critical factor as a small error in grid cell elevation can result in incorrect model results (Vaze, Teng, & Spencer, 2010). First, the raw SRTM DEM was subtracted by -0.25m increments between 0 and 5 meters creating 20 adjusted DEM grids. At each adjusted DEM increment, flood depth and extent were derived using the same 15 dates of gauge data that had matching temporal resolution with the 15 LS tiles. The difference in surface water area (km2) derived by gauge data and Landsat imagery at each DEM increment were then tested using statistical models. A linear regression model, spline interpolation model, 2nd order polynomial model, and 3rd order polynomial model were fit amongst data to observe how flood extent varied with each succeeding DEM increment, where the y-axis represented the DEM adjustment level from 0-5m and the x-axis represented the difference between gauge derived and LS derived flood extent in km2. Theoretically, the perfect level of adjustment would fall along each models fit line at the x=0 plane. Models that best fit the data points were chosen and used for calculating the optimal DEM adjustment level. Once the best fitting models were identified for each tile, the y-value of the fit line at the x=0 plane was recorded. The optimal DEM adjustment level was then derived by computing the average y-value of each tiles best fitting model at the x=0 plane (Supplementary materials Appendix 2).

Gauge derived flood depth and extent were then calculated a final time using the optimal DEM adjustment level with results noted in Supplementary Material Table 2. Lastly, to validate the optimal level of DEM adjustment in calculating gauge derived flood depth and extent, a Student’s t-test was run between Landsat derived flood extent and gauge derived flood extent.

Initial results from a Student’s t-test run between flood extent calculations derived by Landsat imagery classification and gauge data using an unadjusted DEM suggested that there was a statistically significant difference among the outputs (t = 2.6328, df = 28, p-value = 0.01363). Results from a Student’s t-test run between flood extent calculations derived by Landsat imagery classification and gauge data using a -3 meter adjusted DEM suggested that there was no statistically significant difference among the outputs (t = -0.6786, df = 28, p-value = 0.503). Supplementary MaterialFigure 2 displays the extent of surface water area (km2) derived by Landsat imagery and gauge data using a DEM adjusted by -3m in the Sylhet Basin region of Bangladesh.