Researchers from Ohio State University have used an artificial intelligence (AI) algorithm to detect the Vietnam War-era bomb craters in Cambodia from satellite images. They hope that this research will advance the efforts to find and demine unexploded bombs. The model, when combined with declassified U.S. military records, suggests that 44 to 50 percent of bombs in the area may remain unexploded.

Bomb craters on Earth have similarities to meteor craters on the moon. (Source: Ohio State University)

Demining is the search and removal of potentially live bombs and land mines. Methods currently used are not as effective as they need to be. Since the bombing of Cambodia, 64,000 people have died or been injured by unexploded bombs and today, about one person per week is hurt by them. Much of the land covered in the study is used for agriculture, which means farmers have a higher risk of encountering unexploded bombs.

The team started with commercial satellite images of a 100 square kilometer area in Kampong Trabaek, Cambodia. This area was a target of carpet bombing by the U.S. Air Force from May 1970 to August 1973. They used machine learning to analyze satellite images for evidence of bomb craters.

The team knew how many bombs were dropped in an area and the general location of where they fell by looking at craters. Craters indicate where a bomb fell and exploded. After analyzing craters, the team could then determine how many unexploded bombs are left and where they may be.

The study had two stages. In stage one, the team used an algorithm that detects meteor craters on the moon and planets. A human coder found 177 true bomb craters. The algorithm’s results were compared to this number. This algorithm identified 89 percent of true craters with 1,142 false positives. This algorithm was good, but not good enough because bombs can create craters smaller than those made by meteors. Also, many of the existing bomb craters have eroded or been covered by vegetation in the years since the war.

For stage two, the team built up the algorithm to understand the intricacies of how bomb meteor craters are different, including novel features like shapes, colors, textures and sizes. This adjustment eliminated 96 percent of false positives and lost five real bomb craters, with an accuracy of 86 percent.

The newly proposed method increased bomb detection by more than 160 percent. The team analyzed declassified military data that shows 3,205 carpet bombs were dropped in the area. The study suggests that there are 1,405 to 1,618 carpet bombs that are still unaccounted for.

The study was published in PLOS One.