21. Bathy water masking

For shallow water bathymetry (Section 8.31) it is important to distinguish land pixels from water pixels. This allows the algorithm to focus on underwater terrain while avoiding false depth estimates from land features.

A simple and commonly used approach is to threshold the near-infrared (NIR) band 7, where water typically appears darker than land. This method is described in Section 8.31.3.

In complex coastal environments with vegetation, shadows, turbid water, or shallow clear water, the NIR band alone may not provide sufficient separation. This section describes alternative spectral indices that combine multiple bands that may improve land-water discrimination.

21.1. Multispectral image bands

WorldView satellites capture multispectral imagery with eight bands.

Table 21.1 WorldView-3 multispectral bands

Band	Name
1	Coastal
2	Blue
3	Green
4	Yellow
5	Red
6	Red Edge
7	NIR1 (Near-infrared 1)
8	NIR2 (Near-infrared 2)

Other vendors provide similar products.

Individual bands can be extracted from a multispectral image with gdal_translate (Section 16.25). For example, run this for the green band (band 3):

b=3
gdal_translate -b ${b} -a_nodata 0                    \
  -co compress=lzw -co TILED=yes                      \
  -co BLOCKXSIZE=256 -co BLOCKYSIZE=256               \
  input.TIF input_b${b}.tif

The compression and tiling options help with the performance of ASP processing later.

The option -a_nodata 0 sets out-of-footprint pixels to nodata, to avoid division by zero when computing the water indices below.

21.2. Water indices for land-water masking

The following indices provide alternatives to band 7 (NIR1), as described in Section 8.31.3.

21.2.1. NDWI (Normalized Difference Water Index)

NDWI is computed as:

\[\text{NDWI} = \frac{\text{Green} - \text{NIR}}{\text{Green} + \text{NIR}}\]

This index enhances the contrast between water and land. Water typically has positive values, while land and vegetation have negative values. It is effective for general water delineation and separates water from soil and terrestrial vegetation well.

To compute NDWI using image_calc (Section 16.33), do:

image_calc -c "(var_0 - var_1) / (var_0 + var_1)" \
  --output-nodata-value -1e+6                     \
  input_b3.tif input_b7.tif -o ndwi.tif

where input_b3.tif is the green band and input_b7.tif is the NIR band. The --output-nodata-value option sets an explicit nodata value on the output (well outside the valid [-1, 1] NDWI range).

21.2.2. RNDVI (Reversed Normalized Difference Vegetation Index)

RNDVI is computed as:

\[\text{RNDVI} = \frac{\text{Red} - \text{NIR}}{\text{Red} + \text{NIR}}\]

This is the inverse of the standard NDVI (which is high for vegetation). In RNDVI, water appears bright (high values) while vegetation appears very dark (low or negative values). This index is particularly effective in areas with heavy vegetation along the shoreline, such as mangroves or dense forests, where standard NIR masking may be ambiguous.

To compute RNDVI using image_calc run:

image_calc -c "(var_0 - var_1) / (var_0 + var_1)" \
  --output-nodata-value -1e+6                     \
  input_b5.tif input_b7.tif -o rndvi.tif

where input_b5.tif is the red band and input_b7.tif is the NIR band.

21.2.3. OSI (Ocean/Sea Index)

OSI is computed as:

\[\text{OSI} = \frac{\text{Green} + \text{Red}}{\text{Blue}}\]

This index uses the ratio of longer visible wavelengths to blue. It can be useful for specific water conditions, though it is often less robust in clear shallow water than NDWI or RNDVI.

The image_calc command is:

image_calc -c "(var_0 + var_1) / var_2" \
  --output-nodata-value -1e+6           \
  input_b3.tif input_b5.tif input_b2.tif -o osi.tif

where input_b3.tif is the green band, input_b5.tif is the red band, and input_b2.tif is the blue band.

21.3. Thresholding

The resulting index images (ndwi.tif, rndvi.tif, osi.tif) can be converted to binary water masks. This requires computing an appropriate threshold. Here’s an example that invokes otsu_threshold (Section 16.47) for that purpose, with ndwi.tif:

otsu_threshold ndwi.tif

This will print the computed threshold to standard output. This value should then be used in the masking command below.

21.4. Mask creation

When creating binary masks from these indices it is important to note the following.

Polarity reversal: Unlike the raw NIR band (band 7) in Section 8.31.3, where water is darker than land, the spectral indices NDWI and RNDVI make water appear brighter. This affects how binary masks are created from thresholds.

The mask convention is that land pixels have value 1 (or positive values) and water pixels have value 0 (or nodata).

For the NIR band, water pixels are at or below the threshold and land pixels are strictly above, so the masking command is:

threshold=225
image_calc -c "gt(var_0, $threshold, 1, 0)" \
 input_b7.tif -o land_mask.tif

For NDWI and RNDVI, water pixels are at or above the threshold and land pixels are strictly below:

threshold=0.38
image_calc -c "lt(var_0, $threshold, 1, 0)" \
 ndwi.tif -o land_mask.tif

Note the reversed comparison operator (gt vs lt) to maintain the convention that land=1 and water=0.

Site-specific performance: The effectiveness of these indices varies depending on local water conditions, bottom composition, turbidity, and shoreline vegetation.

In testing, NDWI and NIR band 7 showed the most consistent thresholds across images. RNDVI might be effective for vegetated shorelines but could require site-specific threshold adjustment. The results with OSI were not as promising in our experiments (for Key West, FL).

It is recommended to compute all indices and visually inspect the produced masks in stereo_gui (Section 16.71) before selecting the most appropriate one for your specific site.