With the probability of a tower computed for each tile, the next step was to compile these results into a GeoJSON and provide this as a map overlay that the mappers could use on top of their satellite imagery. Therefore, we selected a hard threshold for the probability score; any image with a score at or above this limit would be designated as having a HV tower by the model and incorporated into the GeoJSON overlay for our Data Team to use during mapping. To choose this threshold, we computed the Receiver Operator Characteristic (ROC) curve and plotted the distributions of scores for positive and negative examples (i.e., containing or not containing HV towers). These tiles were included in a GeoJSON map that simply marked a square outlining every positive prediction. The threshold was set at 0.97 in an attempt to balance true and false positives. As in the previous inference step, these GeoJSON overlays were computed in batches according to the X-indices of tiles. Finally, all overlays were concatenated into one file using geojson-merge and uploaded to S3 for the mapping team to download and use within JOSM (Java OpenStreetMap; an OSM editor).