In this paper, multiple instance learning (MIL) algorithms to automatically perform root detection and segmen-tation in minirhizotron imagery using only image-level labels are proposed. Root and soil characteristics vary from location to location, thus, supervised machine learning approaches that are trained with local data provide the best ability to identify and segment roots in minirhizotron imagery. However, labeling roots for training data (or otherwise) is an extremely tedious and time-consuming task. This paper aims to address this problem by labeling data at the image level (rather than the individual root or root pixel level) and train algorithms to perform individual root pixel level segmentation using MIL strategies. Three MIL methods (MI-ACE, miSVM, MIForests) were applied to root detection and compared to non-MIL approches. The results show that MIL methods improve root segmentation in challenging minirhizotron imagery and reduce the labeling burden. In our results, miSVM outperformed other methods. The MI-ACE algorithm was a close second with an added advantage that it learned an interpretable root signature which identified the traits used to distinguish roots from soil and did not require parameter selection. Note to Practitioners-Minirhizotrons provide an efficient and non-destructive way to collect plant roots for studying root system dynamically. However existing software used to extract roots from minirhizotron image require significant, tedious manual marking of roots and soil in the collected imagery. Due to this slow manual process, the ability to collect useful information from a large number of minirhizotron images is bottlenecked. In this paper, we propose an automated approach to segment roots from minirhizotron images. The proposed methods not only automatically identifies and segments root pixels in imagery, but also allow for an efficient approach to label training data. This allows one to be able to retrain the models for adaptation to new environments and soil conditions. The methods we proposed in the paper only require one to label each training image as having roots or not (as opposed to labeling individual pixels).
translated by 谷歌翻译