Reduce population density false positives

src/main/java/com/conveyal/gtfs/error/SuspectStopLocationError.java

@@ -19,7 +19,8 @@ public SuspectStopLocationError(String stopId, long line) {
 
     @Override public String getMessage() {
         return String.format(
-            "Stop with ID %s is located in a 1/4 degree cell with less than 5 inhabitants per square km.",
+            "Stop with ID %s is in a sparsely populated area (fewer than 5 inhabitants per square km in any " +
+                    "neighboring 1/4 degree cell)",
             affectedEntityId
         );
     }

src/main/java/com/conveyal/gtfs/storage/BooleanAsciiGrid.java

@@ -51,19 +51,22 @@ public BooleanAsciiGrid (InputStream inputStream, boolean gzipped) {
     }
 
     /**
-     * Get a grid for places where population density is over 5 people per square kilometer.
+     * Get a grid for places where population density is over 5 people per square kilometer in any neighboring cell.
      * We use the Gridded Population of the World v3 data set for 2015 UN-adjusted population density.
      * This data set was downloaded at 1/4 degree resolution in ESRI ASCII grid format. The grid file was edited
      * manually to eliminate the no-data value header, since I could not find a way to operate on no-value cells in the
      * QGIS raster calculator. Then the raster calculator in QGIS was used with the formula ("glds00ag15@1" > 5),
      * which makes all cells with population density above the threshold have a value of one,
-     * and all others a value of zero (since the no data value in the grid is -9999). This was then exported as another
-     * ASCII grid file, which zips well. The license for this data set is Creative Commons Attribution.
+     * and all others a value of zero (since the no data value in the grid is -9999). Next, the GRASS r.neighbors tool
+     * in QGIS was used to blur this layer, adding to each cell the sum of the neighboring eight cells ("neighborhood
+     * size" of 3, apparently the dimension in each direction). A final logical operation ("blurred@1" > 0) was
+     * applied, with the result exported as another ASCII grid file, which zips well. The license for this data set
+     * is Creative Commons Attribution.
      * See http://sedac.ciesin.columbia.edu/data/collection/gpw-v3
      */
     public static BooleanAsciiGrid forEarthPopulation() {
         try {
-            InputStream gridStream = BooleanAsciiGrid.class.getResourceAsStream("gpwv3-quarter-boolean.asc");
+            InputStream gridStream = BooleanAsciiGrid.class.getResourceAsStream("gpwv3-quarter-buffer-boolean.asc");
             return new BooleanAsciiGrid(gridStream, false);
         } catch (Exception ex) {
             throw new RuntimeException(ex);