Publications Thursday: Needles in the crop mapping haystack

February 13th, 2020

Relatively infrequent crops (for example, field peas in Australia) are hard to map using satellite data. Mapping algorithms that are globally optimal turn out to be error-prone for crops that occupy a small fraction of land area. This “class imbalance” problem can, in turn, disproportionately affect the usefulness and credibility of the resulting maps.

Early-career research Franz Waldner and his colleagues in Digiscape’s Graincast project have developed methods for improving the mapping accuracy of rare crops while having the smallest possible effect on overall mapping accuracy. Their key discovery is that no one  “balancing” algorithm works well in all circumstances. So, they have used an algorithm-selection method (called F-race) that chooses, for a specific map, the right way of trading off improvement in performance for rare crops and all crops.

Their work has been published by Remote Sensing and Environment; the full paper is here