In the last year, we’ve worked on quantifying agricultural yield in Africa using EO satellite imagery. Identifying field boundaries is always the starting point for agricultural studies, and it would be great if we could do it automatically. In America you see agricultural fields that are uniform, cleared, and irrigated; but, in Africa, smallholder fields tend to be none of those things. Looking at ag images from Africa, and thinking about delineating African agricultural field boundaries, will soon have you questioning whether the idea of field boundaries really makes any sense at all when a field might have several different crop types, a few trees, and a huge termite mound in the middle of it.
But land tenure, and whether it can be ascertained on a large scale from remote sensing imagery, is a problem of interest to agencies working on sustainable land development. There are clues as to where the boundaries of fields are: the field treatment may change, or there may be tree lines, paths or rivers. But sometimes it takes a judgment call, and you get the best judgments by asking a lot of people their opinions.
So we recently kicked off a crowdsourcing campaign on the Tomnod crowdsourcing website, using Tomnod’s new polygon Draw capability, to try to solve the problem. We asked the Tomnod crowd to collect field boundary candidates over a small ag scene in Mali, using intuition with only a few guidelines. I drew a bunch of field boundaries for the campaign, and found it a very easy interface, fun and kind of addictive. So far, the crowd has drawn about 6000 field boundary polygons.
But where is this going? Here’s what I’m hoping:
- In the end we’ll have a number of candidates for possible fields, contributed by different individuals. ‘
- We’ll identify the serious field candidates based on a validation campaign, or some sort of consensus measure.
- We’ll aggregate them into a ‘best field’ estimate, in the same way that taking an average of a set of estimates of a number improves the estimate of the number, by reducing error.
- We’ll either use the high-quality results themselves, or we’ll have a rich set of training data for developing a machine learning approach to automated field boundary delineation.