Plant leaf deep semantic segmentation and a novel benchmark dataset for morning glory plant harvesting

Abstract

Computer vision and deep learning have made substantial progress in the areas of agriculture and smart farming, particularly for enhancing crop production using image segmentation techniques for crop yield prediction. Further improvements to crop yield prediction results can be achieved by developing accurate and efficient methods. In response to such demands, this paper proposes a novel convolutional neural network architecture, called densely connected SegNet (D-SegNet) and demonstrates its advantages on plant segmentation using a new morning glory plant dataset, and also on a complimentary publicly available dataset to promote research in this direction. The D-SegNet is evaluated using 10-fold cross validation. It achieves performance better than the state-of-the-art SegNet algorithm. The evaluated precision, recall and F1-score values are 98.20%, 90.64% and 94.26%, respectively, for the morning glory plant dataset. The intersection over union (IoU) value in the image segmentation tasks is 90.56%. A series of experiments on the morning glory plant dataset as well as on the publicly available dataset were conducted. The results show that the proposed method achieves accurate segmentation results and can be useful for assessing the plant weight during harvesting. In summary, this new plant segmentation network, D-SegNet, could form an important component of future cloud-based machine learning systems to predict crop yield from noisy smartphone images taken in the field.

Plant phenotyping relevance

植物画像のセマンティックセグメンテーション手法とベンチマークデータセットを開発・評価し、収穫時の植物重量推定という植物形質推定への利用を示しており、方法が中心である。

Code and data availability

The paper claims the morning glory dataset was 'made public' on GitHub (refs [30,31]), but the Data availability statement contradicts this: 'The data that has been used is confidential.' The cited GitHub links are generic profile pages (not specific dataset deposits), are not among the allowed_urls, and no author code

No evidence-backed public reproduction asset is currently recorded.