Overview and Data Sources

This methodology has been refined and simplified into the fewest possible number of operations to accomplish the objective of detecting large vessels within an area of interest. The simplified methodology still achieves high accuracy and precision, as detailed in the following section. This methodology uses Sentinel-1 Ground Range Detected (GRD) SAR imagery in vertical transmit, horizontal receive (VH) polarization acquired in Interferometric Wide (IW) mode. The specific boundaries of vessel grouping areas around each of the three major port terminals were defined by the researcher as vector polygons.

Analytical Process

The analytical process for quantifying maritime activity within an area of interest is encapsulated in a four-step approach. The following procedures were executed using JavaScript on an initial dataset of all Sentinel-1 images captured since 2015. All of these procedures were performed in a single Google Earth Engine script.

Image Collection Filtering

The collection of all Sentinel-1 GRD IW VH images is filtered to a set of 494 valid images that a) were captured between 1/1/2015 - 9/1/2025; b) intersect an area of interest point at 38.6949°N, 125.3208°E; and c) were not captured during the months of January or February. The exclusion of certain months is applied due to overdetection during periods of heavy ice on the Taedong River, which is discussed further in the Analytic Confidence section.

Binary Mask Generation and Clipping

For each image in the filtered collection, a binary mask is generated where pixels with VH backscatter values >= -20 dB are assigned values of 1 and all other pixels are assigned values of 0. This binary mask is then clipped to the extent of the river itself using a manually created polygon of the water extent. This step produces a binary image of high backscatter clusters on the river for every image.

Cluster Vectorization and Filtering

For each binary image, every contiguous cluster of high backscatter pixels is reduced to a unique vector object. Vectors that have total sizes smaller than 15px or larger than 200px are filtered out. This controls for most artifacts and removes any small vessels outside the scope of the analysis. This step produces a set of likely ship vector objects for every image.

Terminal Grouping

Based on the researcher’s expertise and long-term pattern observation of ship clustering locations, approximate polygons were developed around each of the three main terminals to include areas of the river in which any vessels were likely approaching or departing the given terminal. Separate large vessel vector counts were then generated for every image for each of the following sub-areas:

• Approach to the West Sea Barrage
• Entire River Adjacent to Nampo Port
• POL Terminal Proximal Area
• Bulk Goods and Coal Terminal Proximal Area
• Container Terminal Proximal Area
• Upriver Areas (From Nampo to Chollima)