Machine learning aside, there is room for improvement on the human component of our mapping efforts carried out by the Data Team. The concept of using machine learning to guide mapping was new and untested for our professional mappers and our ML team, so we collected ideas to streamline the process throughout this project. Our Data Team found that navigating through the ML model's predictions was tedious — initially they were repeatedly zooming in to map a few towers and then zooming out to reestablish a high-level view of the predictions. This constant reorientation of the view also made it difficult to keep track of which areas had been reviewed. As discussed in the Methodology Section, the Data Team built its own To-Fix plugin within JOSM so they could click through the model's predictions with a single button. In the future, we will also explore the potential to "jump" several towers at a time when mapping if only the HV lines are of interest. This could reduce mapping efforts up to 30% in areas where multiple towers in a row are visible.
Another limitation was that the ML predictions were not available to view in the tasking manager -- the piece of software used to organize a group of mappers working in the same geospatial area. Initially, there was no way to view the ML predictions in this overlay, but we were able to create the ability to add the ML overlay just like other map layers.
The Data Team also noted that image quality varied across entire countries. In some cases, it was difficult to accurately trace HV infrastructure due to blurry or otherwise poor-quality imagery. Future efforts should explore other sources of satellite imagery that are higher quality and up to date. An engineering consult who is familiar with typical HV network construction might also be advantageous. Someone with domain knowledge should be able to rule out ambiguous features in the imagery; examples include connections that appear to involve three HV towers, changes in the size (and possibly operating voltage) of HV towers, or areas where the HV lines appeared to bifurcate. Areas where many different HV lines came together (often near substations in cities) were also difficult to delineate as were towers in complicated urban environments. These special cases were relatively rare, but they are difficult to map from overhead imagery at zoom 18 resolution.