"Pumpkins are in peak demand this time of year, and farmers need to provide an accurate assessment of crop counts to retailers (jack-o-lanterns) and processors (canned pumpkins). Harvest is also the time of year when farmers measure yields and make management decisions that will lower costs and increase yields in the following year."
"Drones are generating more data than we know what to do with. Every mission can generate thousands of images (or hours of video) that need analysis. With machine learning, we can teach algorithms to detect something we care about (e.g., leaf, soil, and pumpkin), classify it, and localize it in space. From there, we can do further processing, such as measuring the diameter and mapping the density. By incorporating additional data streams and decision-making capabilities, your machine learning becomes artificial intelligence (AI). Fortunately, there is still a role for humans. Raptor Maps fed the machine learning algorithm labeled examples of pumpkins (“positives”), as well as examples of vines and soil (“negatives”). The negative examples are important, since reflections on large, broad leaves would otherwise be flagged by the algorithm as “pumpkins” and result in an artificially higher count. As you can see in the map below, the pumpkins may occasionally be touching, so Raptor Maps had to ensure that the software could smartly identify this area as two separate pumpkins, instead of one large one. Each individual crop is given a unique identifier. This makes it easier to monitor and track individual crops during repeat inspections, even if they slightly shift over time. This type of approach brings the industry closer to analytics and management at the single-crop level. Once each pumpkin is identified, the yield map is created by dividing the field into pixels (in this case 1/32 acres). The resulting yield maps are shapefiles that are compatible with farm management software, such as the John Deere Operations Center, Case IH AFS, Climate FieldView, and FarmLogs."
"Raptor Maps fed the machine learning algorithm labeled examples of pumpkins (“positives”), as well as examples of vines and soil (“negatives”). The negative examples are important, since reflections on large, broad leaves would otherwise be flagged by the algorithm as “pumpkins” and result in an artificially higher count. Pumpkins may occasionally be touching, so Raptor Maps had to ensure that the software could smartly identify this area as two separate pumpkins, instead of one large one. Moreover, Raptor Maps used a neural network."
Images or videos captured by drones
Proof of concept; results not yet available