J2.4
Weather Modelling to Enable Precision Forecasting for Agricultural Operations

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Monday, 5 January 2015: 4:45 PM
131C (Phoenix Convention Center - West and North Buildings)
Lloyd A. Treinish, IBM Research, Yorktown Heights,, NY; and J. P. Cipriani, A. Praino, C. Cox, and D. Reckford

The operation of any farm is dependent to a significant degree upon weather conditions, for both routine conditions and relative extremes in wind, precipitation or temperature. In the US alone, approximately, $15B of the agricultural industry is sensitive to the variability in weather each year. For example, planning for the applications of pesticides, herbicides or fertilizer, the scheduling of irrigation, planting and harvesting, and field preparation are impacted by the timing and location of various weather characteristics, such as the type, amount and rate of precipitation, temperature, humidity, solar radiation and/or wind. The interaction of the lower boundary layer with the land surface leads to other effects that should be considered such as soil moisture, runoff and temperature, and evapotranspiration. Hence, with precipitation events, local topography and weather conditions influence water runoff and infiltration, which directly influence the effectiveness of irrigation, or the environmental impact on nearby land and water bodies, among other farm operations. In addition, there are specific weather conditions associated with specialized operations such as determining when and where water should be sprayed in citrus groves to shield fruit from damage due to temperatures falling below the freezing point of water. There are also indirect considerations, such as interruption in the supply chain for farming operations due to weather conditions. All of these cases illustrate the need for more precise and accurate information to enable a farmer to plan operations in advance with confidence to improve crop yield.

Therefore, the availability of highly localized weather model predictions focused on the local conditions of a farm has the potential to mitigate the impact of uncertainty in weather conditions on agricultural operations. Typically, predictive information at such a scale is simply not available. Hence, what optimization that is applied to these processes for scheduling and resource allocation to enable proactive operations utilizes either historical weather data as a predictor of trends or the results of continental-scale weather models. Neither source of information is appropriately matched to the temporal or spatial scale of a farm. While the deployment of sophisticated environmental instrumentation on farms has expanded in recent years, they can only provide near-real-time assessment of observations of current weather conditions. They may have the appropriate geographic locality, but by their very nature, such data are only directly suitable for a reactive response.

To address this situation in many industries, IBM Research has developed a service, dubbed “Deep Thunder”, which has the ability to predict specific events or combination of weather conditions with sufficient spatial and temporal precision, and lead time to enable proactive allocation and deployment of resources (people and equipment) to increase time for more reliable planning. In particular, the Deep Thunder service provides local, high-resolution weather predictions customized to business applications for weather-sensitive operations from one to three days ahead of time.

To begin to explore the application of these ideas to enable precision agriculture, Deep Thunder has been adapted to the Lower Flint River Basin in southwestern Georgia, working with The Nature Conservancy, Flint River Soil and Water Conservation District, the Natural Resources Conservation Service of the US Department of Agriculture, the University of Georgia, and a consortium of local farmers. This region is the leader in the production of peanuts and a major producer of cotton and sweet corn. Efficient production of these crops is an intensive user of water, which has been further challenged in recent years due to drought. Hence, there has already been investment in the region with deployment of environmental instrumentation and variable-rate irrigation systems.

IBM Research has developed a multi-resolution forecasting domain for the geographic area including the Lower Flint River Basin and the surrounding region based upon a highly specialized configuration of the WRF-ARW (version 3.5.1) community model. This included an analysis of available data, and determination of the appropriate resolution (e.g., 1.5 km horizontal resolution for the inner nest and 4.5 km for the surrounding area with sufficient vertical layers at key altitudes to handle terrain that influences weather in the region). The result was an operational capability with 72-hour forecasts, which are updated every 12 hours and have up to 66 hours of lead time, depending on the client requirements and available computational resources. It incorporates a diversity of input data sets ranging from NCEP RAP for background conditions, NCEP NAM for lateral boundaries, three-dimensional variational data assimilation from several thousand surface and near-surface observation stations operated by Earth Networks, NOAA and other agencies, and surface conditions (soil, land use, vegetation, terrain and SSTs) derived from remotely sensed observations provided by NASA and USGS. The surface data are particularly important for stable and accurate boundary layer and land surface representation.

In addition to background on the meteorological and business issues for agriculture, we will discuss the on-going work, the overall approach to the problem, some specifics of the solution, and lessons that were learned through the development and deployment. This will include a discussion of the analysis and the results of operational forecasting that began in late 2013. In addition to the evaluation, we will present how the forecasts are being used, including the deployment of customized visualizations. We will also discuss the overall effectiveness of our particular approach for these and related applications, issues such as calibration of data and quantifying uncertainty as well as recommendations for future work.