## Water Mask Mosaic

The mosaic was generated from a collection of per-scene water masks (by product) that were originally produced as a preprocessing step for the first pan-Arctic [IWP dataset](https://arcticdata.io/catalog/view/doi:10.18739/A2KW57K57). The mosaicking workflow was adapted from the Polar Geospatial Center (PGC) [imagery utilities](https://github.com/PolarGeospatialCenter/imagery_utils). We modified the scene-selection logic to better handle overlapping masks and to produce a smoother, more consistent mosaic across the full collection using a combination of exposure, acquisition month and year. [Updated water mosaic code with HPC helper scripts](https://github.com/UConnARIAL/WaterMaskMosaic)

## Geopackage generation (.gpkg)
Before the generation of the final gpkg with 2 layers
(1) Original water raster, 
(2) Vector layer with area in m^2 computed for each water body after cleaning objects <2m^2
For the calculation the layer was projected to: EPSG:6931 — WGS 84 / NSIDC EASE-Grid 2.0 North. Which is a Lambert Azimuthal Equal Area projection for the Northern Hemisphere
 
We also cleaned the layers using the NSIDC water polygons, applying a 500 m buffer along the ocean boundary. Some ocean areas near the boundary may still be retained as water due to boundary generalization and polygonization effects.

### Water-mask generation (NIR + median smoothing + Otsu threshold)
Water masks were generated automatically from each input Maxar Scenes using a lightweight spectral segmentation workflow designed to exclude open-water areas from downstream processing (iwp inferencing). The approach leverages the near-infrared (NIR) band, where liquid water typically exhibits very low reflectance relative to most land surfaces, enabling a strong bimodality in the NIR histogram for many scenes[1][2].

For each image, the pipeline: 
(1) converts the input Maxar Images to an 8-bit representation for efficient processing (GDAL translate), 
(2) extracts the NIR band, 
(3) applies a rank-based median filter with a circular structuring element (disk radius = 5 pixels) to suppress speckle and small artifacts, 
(4) applies additional Gaussian smoothing (σ = 2.0), and 
(5) computes a global Otsu threshold to separate two dominant intensity classes. 
Pixels are then binarized using this threshold and written to a 1-bit GeoTIFF (NBITS=1) while preserving the original georeferencing (GeoTransform and projection). 

Otsu’s method provides an unsupervised, per-scene threshold selection, and median filtering is a standard robust denoising step prior to global thresholding.

The [code used](https://github.com/PermafrostDiscoveryGateway/MAPLE_v3/blob/main/maple_workflow.py) to genreate the water mask as part of the overall MAPLE workflow [3] is avaialble with the PDG GitHub repositories.

[1] N. Otsu, “A Threshold Selection Method from Gray-Level Histograms,” IEEE Trans. Syst., Man, Cybern., vol. 9, no. 1, pp. 62–66, 1979, doi: 10.1109/TSMC.1979.4310076.