SpikeletFCN: Counting Spikelets from Infield Wheat Crop Images Using Fully Convolutional Networks

Abstract

Currently, crop management through automatic monitoring is growing momentum, but presents various challenges. One key challenge is to quantify yield traits from images captured automatically. Wheat is one of the three major crops in the world with a total demand expected to exceed 850 million tons by 2050. In this paper we attempt estimation of wheat spikelets from high-definition RGB infield images using a fully convolutional model. We propose also the use of transfer learning and segmentation to improve the model. We report cross validated Mean Absolute Error (MAE) and Mean Square Error (MSE) of 53.0, 71.2 respectively on 15 real field images. We produce visualisations which show the good fit of our model to the task. We also concluded that both transfer learning and segmentation lead to a very positive impact for CNN-based models, reducing error by up to 89%, when extracting key traits such as wheat spikelet counts.

Unsupervised Domain Adaptation For Plant Organ Counting

“…Even though their results show high accuracy, counting spikelets from real field images requires further fine-tuning of their model. Alkhudaydi et al [2] proposed a method to count spikelets from infield images using a fully convolutional network named SpikeletFCN. They compared the performance of training SpikeletFCN from scratch on infield images and fine-tuning the model, which was initially trained using images taken in a controlled setting [33].…”

“…For the wheat spikelet counting task, we used the general problem definition of wheat spikelet counting as stated in [2]. Under this setting, our second experiment tries to adapt the wheat spikelet counting task from images taken in a greenhouse to images captured in an outdoor field.…”

“…Object counting is a relevant task in image-based plant phenotyping, notably for counting plants in the field to estimate the rate of seedling emergence, and counting plant organs to estimate traits relevant for selection in crop breeding programs. For example, counting spikes or heads in cereal crops is a relevant trait for estimating grain yield [33,27,2], counting flowers for estimating the start and duration of flowering is relevant for demarcating plant growth stages [23,3,45], and counting leaves and tillers is a relevant trait to assess plant health [16,1]. Leaf counting, in particular, has been a seminal plant phenotyping task, thanks to the CVPPP leaf counting competition dataset [30,29].…”

“…We evaluated our proposed domain adaptation method on two object counting tasks (wheat spikelet counting and rosette leaf counting) each with a different challenge. The wheat spikelet counting task adapts from indoor images to outdoor images and presents challenges due to self-similarity, self-occlusion, and appearance variation [33,2]. The leaf counting task adapts from one plant species to a different plant species and presents variability in leaf shape and size, as well as overlapping and occluded leaves [19].…”

“…CropQuant dataset [47]: This dataset is composed of images collected by the Norwich Research Park (NRP). As a ground truth, we used the dot annotations created by [2], where they annotated 15 of the images from the CropQuant dataset. The ground truth has a total of 63,006 spikelets.…”