Use of advanced modelling methods to predict dothistroma needle blight on Pinus radiata at a fine resolution within New Zealand

Abstract

Dothistroma needle blight (DNB) is a damaging disease of many Pinaceae host taxa including the widely grown species Pinus radiata D. Don. However, little research has modelled the distribution of this disease at a fine spatial resolution from geospatial surfaces using advanced non-parametric methods. This study uses an extensive national dataset (n = 6276 observations) that describes the severity of DNB on plantation grown P. radiata with measurements expressed as a percentage of foliage within the stand affected by the disease (range 0–90%). Using these data the objectives of this research were to (i) compare the predictive precision of models of disease severity constructed using parametric and non-parametric approaches, (ii) determine whether the use of regression kriging is able to improve model fit, (iii) identify the key environmental determinants of disease severity using the most precise modelling methods and (iv) develop a fine scale map showing variation in disease severity throughout New Zealand. All predictions were made on a test dataset (n = 1883) that was not used for model fitting.Disease severity was most precisely predicted by the two non-parametric models eXtreme Gradient Boosting (RMSE = 11.37%, R2 = 0.231) and random forest (RMSE = 11.28%, R2 = 0.246) which both markedly outperformed the parametric multiple regression model (RMSE = 12.17%, R2 = 0.120). The use of regression kriging considerably increased the precision for the multiple regression (RMSE = 11.88%, R2 = 0.161) but did not improve model precision for the two non-parametric models.Random forest was the most precise model and included 14 variables, of which the time of the observation, stand age, relative humidity and solar radiation had the strongest influence on disease severity. A fine resolution map (25 m) created using random forest showed disease severity to be highest in northern and western areas of New Zealand, where relative humidity and rainfall are high. Low disease severity was predicted on the drier east coast of New Zealand and values were particularly low in the far south, where mean annual solar radiation is relatively low. This surface will assist forest managers in narrowing the search radius for the disease and prioritising areas for targeted deployment of disease resistance clones, provenances or alternative species.