Weed Infestation Identification Using Hierarchical Crowdsourcing

Weed infestation is a common problem in agriculture that adversely affects crop production. Given severe constraints on the budget of many landgrant universities due to the economic downturn, outreach services have taken a hit. To adapt to current economic climate without adversely affecting the quality of outreach program for weed management, we present a hierarchical system that uses image captured by smartphone, a backend image processing algorithm, and two levels of crowdsourcing approaches to identify weed images. The first of the two crowdsourcing levels consist of non-expert crowd contributed by Amazon Mechanical Turk and the second level consisting of expert crowd comprising of county extension agents. A probabilistic decision engine was used, in an unsupervised manner, to determine the suitability of two levels of crowdsourcing approaches for identifying the weed image. The designed system was found to have low latency with high accuracy for identifying weed from captured images. The designed system accurately identified test weed within 3 hours of its submission using minimal human intervention. The system has shown good potential to support weed management related outreach programs in land-grant universities.