Counting spikelets from infield wheat crop images using fully convolutional networks

Abstract

Wheat is one of the world’s three main crops, with global consumption projected to reach more than 850 million tons by 2050. Stabilising yield and quality of wheat cultivation is a major issue. With the use of remote sensing and non-invasive imaging technology, the Internet of things (IoT) has allowed us to constantly monitor crop development in agriculture. The output of such technologies may be analysed using machine-learning algorithms and image processing methods to extract useful information for crop management assistance. Counting wheat spikelets from infield images is considered one of the challenges related to estimating yield traits of wheat crops. For this challenging problem, we propose a density estimation approach related to crowd counting, SpikeCount. Our proposed counting methods are based on deep learning architectures as those have the advantage of being able to identify features automatically. Annotation of images with the ground truth is required for machine learning approaches, but those are expensive in terms of time and resources. We use transfer Learning in both tasks, segmentation and counting. Our results indicate the segmentation is beneficial as focusing only on the regions of interest improves counting accuracy in most scenarios. In particular, transfer learning from similar images produced excellent results for the counting task for most of the stages of wheat development.