Counting Strawberry Flowers On Drone Imagery With A Sequential Convolutional Neural Network

Strawberry cultivation is notoriously unpredictable due to inherent fluctuating yields. The possibility for yield prediction has useful and important operational and logistical advantages. One possible indicator for yield prediction is the number of blooming flowers. Counting flowers however is very labor intensive and difficult to extrapolate over larger areas due to local variations. In this paper we present an automated predictive strategy for counting strawberry flowers in agricultural settings by image processing of drone imagery. The counting algorithm is based upon density prediction with a convolutional neural network (CNN). Furthermore, we present the data collection, annotation and augmentation, and focus on several practical issues that can have a large effect on the performance of this technique. The automated counts are assessed against manually annotated results which differ by less than 10%, and is comparable to the variation in annotation among different observers.