Late Blight Prediction and Analysis

Abstract: Phytophthora infestans cause severe losses in potato cultivation. Currently one of the most effective ways to reduce the related loss is using fungicides. For these to have optimal effect they must be applied approximately 10 days before the first visual sign of late blight. As the appearance of late blight is hard to predict there is heavy overcompensation of fungicides nowadays. By better forecasts less chemical compounds will be used lowering the environmental impact. It may also increase the profit from potato cultivation as less money goes to fungicides. There are several models trying to predict the late blight appearance date of which one of the most well known is SIMCAST. Also Cox regression models have recently been looked into. Is it possible to improve these models? Moreover what financial impact will it have for farmers? First an empirical investigation of the data was conducted for a general understanding of the subject. Then it was analyzed using, linear regression, Cox regression, Elastic net and cross-validation. To fill in missing data imputation was used. The best performing model was obtained through Cox regression using SIMCAST and planting day as covariates. It had approximately the same failure ratio as foliage prediction which is currently used, but had 23% more accurate predictions and lower standard deviation. SIMCAST was principally out-performed by all other models. Using elastic net it was possible to obtain models which was approximately as good as the best model. Using the best model savings around 422 SEK/ha can be made from reduced fungicides alone. The newly obtained models has potential to increase the profit of potato cultivation. Also by these the accuracy of prediction would contribute to lower environmental impact. Notable is that the data used is not optimal for the experiment as some had to be estimated etc. hence before relying on these models field experiments should be considered.