Improving the prediction accuracy of Thrips tabaci activity using deep learning modelling

The onion thrips, Thrips tabaci Lindeman, is a widespread pest causing significant economic damage. In Flanders, T. tabaci reduces the yield, quality and marketability of a wide range of open field vegetables, including leek, onion, Brussels sprouts and head cabbage. Its short life cycle and ability to rapidly become resistant to insecticides make T. tabaci difficult to control. There is a need for an integrated control strategy, which involves careful and reasoned handling of the remaining authorised chemical agents and focuses on prevention, pest monitoring and adapted cultivation techniques that stimulate natural enemies. Determining the exact timing of initial thrips activity and pest population size is a prerequisite for effective pest management. The T. tabaci activity is highly weather-dependent and can be predicted based on the prevailing abiotic conditions. To date, mainly temperature-based degree-day models have been used for this purpose. However, in practice, these models are not always accurate, which raises questions about their suitability for use in warning systems. Hence, there is a need for more complex prediction models that consider additional abiotic factors, such as precipitation, relative humidity, wind, and light conditions. We assessed how accurately a Recurrent Neural Network (RNN) model, based on a Long Short-Term Memory (LSTM) network, can predict both the timing and intensity of adult thrips activity, based on various weather parameters. Training the model is done using weekly onion thrips observations with blue sticky traps and weather data from several Flemish leek fields of the past seven years. The predictive power is then assessed by allowing the model to predict two years of thrips activity in a field, given only the weather data. In addition, predicted thrips activity based on the RNN-LSTM model is compared with outputs from the currently used degree-day models. The results will enable us to select the most appropriate model for implementation in the current warning system for T. tabaci in open field vegetables.