Convolutional neural networks in predicting cotton yield from images of commercial fields

[Display omitted] •Images of the cotton plants in commercial fields were acquired by a mobile device.•We present an approach robust exploring environmental condition.•The best result for count cotton bolls obtained an accuracy of 8.84%•Predicting cotton yield showed a high precision (R2 = 0.93).•Deep Learning networks were efficient and viable to be used in harvesting machines. One way to improve the quality of mechanized cotton harvesting is to change harvester settings and adjustments throughout the process, according to information obtained during the operation. We believe that yield predictions are important for managing the quality of operation, aiming at increasing efficiency and reducing losses. Therefore, this study aimed to develop an automated system for cotton yield prediction from color images acquired by a simple mobile device. We propose a robust approach to environmental conditions, training detection algorithms with images acquired at different times throughout the day, and evaluating three different scenarios (low-, average-, and high-demand computational resources). The experimental results for the average demand computational scenario, which are suitable for real-time deployment on low-cost devices such as smartphones and other ARM-processed devices, indicated the possibility of counting bolls using images acquired at different times throughout the day, with mean errors of 8.84% (∼5 bolls). Furthermore, we observed a 17.86% error when predicting yield using 205 images from the testing dataset, which is equivalent to about 19.14 g.