Reduced Antibiotic Use with Revised
MARYBLYT Forecasting

Introduction - Fire blight, caused by the bacterium Erwinia amylovora, is one of the most destructive and difficult-to-control diseases of apple. The ability to predict the onset of fire blight has been the most limiting factor in improving the overall management of the disease. MARYBLYT is a computer program for forecasting fire blight and the appearance of symptoms that follow. The forecaster has been used widely throughout the world and it's performance has received mixed reviews.

The objectives of this research are:

1) Revise MARYBLYT to calculate a system of risk points as function of the MARYBLYT risk factors, then using the revisions, evaluate various management-action thresholds based on the accumulation of risk points;

2) Modify MARYBLYT to account for varietal susceptibility, orchard age, and inoculum potential;

3) Develop a user-friendly, windows-based version of MARYBLYT. The modifications will incorporate flexibility into MARYBLYT allowing users to choose suitable action thresholds based on variety, orchard age, inoculum pressure, and their comfort for assuming risk.

Ultimately, these changes will reduce the cost of managing fire blight by emphasizing efficient and minimizing unnecessary applications of antibiotics. This project is funded through the USDA-CSREES Northeast IPM program and is a collaborative effort with Dr. Alan Biggs at West Virginia University and Gary Lightner.

Accomplishments:

A) By the end of 2002, close to 100 historical orchard data sets from locations around North America and the UK were compiled and the predictive accuracy of MARYBLYT (MB) was evaluated using receiver operator characteristic (ROC) curve analysis. Historical data included climatic data during bloom, whether or not disease occurred, and when symptoms appeared. MB uses 4 parameters to predict blossom blight infection: 1) flowers open; 2) epiphytic inoculum potential (EIP) exceeds 100; 3) rain or heavy dew within the last 24 hr; and 4) average temperature greater than 15.6 C. Risk of infection is rated as 'low', 'moderate', 'high', or 'infection' depending upon whether one, two, three or all four of the thresholds have been exceeded, respectively. Although infection is predicted only when all 4 thresholds are exceeded, infections can occur in less favorable conditions, i.e., when only 2 or 3 thresholds are exceeded. In practice, any combination of those 4 thresholds could be used as a rule to trigger a management action. ROC analysis permits us to evaluate each possible combination of thresholds individually as a rule to trigger management action. For each data set, the historical weather data was run through MB and, for each possible rule, it was recorded whether infection was predicted or not, and the prediction was compared to what actually occurred. For each rule, the number of true negatives (no infection predicted and no disease observed), false negatives (no infection predicted and disease observed), false positives (infection predicted and no disease observed) and true positives (infection predicted and disease observed) was recorded. The true positive proportion (TPP) is defined as the number of true positives per number of data sets with disease. The false positive proportion (FPP) is the number of false positives per number of data sets without disease. A perfect rule would have a ratio between TPP and FPP of 1:0. The most reliable predictions were made when decisions were based on having thresholds of all four parameters exceeded. The TPP:FPP ratio was 0.70:0.48. The worst rule, 0.97:0.92, was made when rain and blossom were sole parameters. Many growers act when MB considers the risk of infection as 'high' or 'infection'. The ratio for this rule was 0.88:0.91. Other permutations of the parameters were analyzed but not reported here.

B) The software has been upgraded and converted into a Windows-based format and is currently being tested. Additions include: auto sensing for Fahrenheit or Celsius and a global threshold option that will allow the operator to customize changes. Data is now stored in an ASCII format and can be edited with any text editor software. This eliminates the proprietary data storage structure of the old system.

C) To date, two seasons worth of blossom and shoot blight data have been gathered from 9 commercial orchards in New York. The data represent a wide range of orchard ages and cultivars. Blossom blight data was also collected in West Virginia at a further 27 orchard sites. All blight data was in conjunction with weather data. This data will be used in future analyses to add additional parameters and improve current MB thresholds to improve fire blight forecasts.