Modeling and Monitoring Crop Disease in Developing Countries

Information about the spread of crop disease is vital in developing countries, and as a result the governments of such countries devote scarce resources to gathering such data. Unfortunately, current surveys tend to be slow and expensive, and hence also tend to gather insufficient quantities of data. In this work we describe three general methods for improving the use of survey resources by performing data collection with mobile devices and by directing survey progress through the application of AI techniques. First, we describe a spatial disease density model based on Gaussian process ordinal regression, which offers a better representation of the disease level distribution, as compared to the statistical approaches typically applied. Second, we show how this model can be used to dynamically route survey teams to obtain the most valuable survey possible given a fixed budget. Third, we demonstrate that the diagnosis of plant disease can be automated using images taken by a camera phone, enabling data collection by survey workers with only basic training. We have applied our methods to the specific challenge of viral cassava disease monitoring in Uganda, for which we have implemented a real-time mobile survey system that will soon see practical use. The economies of many developing countries are dominated by an agricultural sector in which small-scale and subsistence farmers are responsible for most production, utilizing relatively low levels of agricultural technology. As a result, disease among staple crops presents a serious risk, with the potential for devastating consequences. It is therefore critical to monitor the spread of crop disease, allowing targeted interventions and foreknowledge of famine risk. Currently, teams of trained agriculturalists are sent to visit areas of cultivation across the country and make assessments of crop health. A combination of factors conspire to make this process expensive, untimely and inadequate, including the scarcity of suitably trained staff, the logistical difficulty of transport, and the time required to coordinate paper reports. Although computers remain a rarity in much of the developing world, the near-ubiquity of mobile telephony has brought low-cost and reliable wireless internet services to broad regions that still lack electricity, running water and paved roads. Among other benefits, the prevalence of mobile computing devices at last offers a feasible alternative to paper-based data gathering. Copyright c © 2011, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. This paper describes three innovations. First, we show how accurate response surface models of disease incidence and prevalence can be built using limited data, collected according to existing survey techniques. One challenge arising from the need to cohere with existing practices is that the levels of disease severity across the spatial field must be expressed as ordinal values, requiring the use of ordinal regression techniques. Second, we show how these models can be used in an active learning framework, determining in real time where survey workers should gather their next samples. This approach has workers gather data non-uniformly in order to maximize the value of the information gathered, as measured by a utility function elicited from domain experts. Because workers follow fixed circuits, our active learning task is an online optimization problem: each field must either be surveyed immediately or passed by. Finally, we present computer vision techniques for using camera-enabled mobile devices to make disease diagnoses directly, allowing reliance on survey workers with lower levels of training, and hence reducing survey costs. Specifically, given expert-annotated images of single cassava leaves, we demonstrate classification based on color and shape information. We have applied these ideas to the domain of viral cassava disease monitoring in Uganda. Cassava is the third largest source of carbohydrates for human consumption worldwide, providing more food calories per cultivated acre than any other staple crop. It is an extremely robust plant which tolerates drought and low quality soil. The foremost cause of yield loss for this crop is viral disease (Otim-Nape, Alicai, and Thresh 2005), a major factor keeping East African farmers trapped in poverty (The Economist, 2011). We have developed a mobile survey system (see screenshot in Figure 3) which is currently being field trialled in partnership with Uganda’s National Crops Resources Research Institute (NACRRI), and expect this to be used in their upcoming crop survey. Source code and survey data are available at http://cropmonitoring.appspot.com. Spatial density estimation In a crop disease survey, each plant is assigned a disease level yi ∈ {d1, ..., dD}, usually by visual inspection of the aerial parts of the plant. A two-class survey might be done, where d1 and d2 represent healthy and diseased plants respectively; though often for cassava, disease levels are assigned categories from d1 (entirely healthy) to d5 (very severe disease, Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence