SegRoot: A high throughput segmentation method for root image analysis

Abstract

The measurement of root growth over time, without destructive excavation from soil is important to understanding the development of plants, communities and ecosystems. However, analyzing root images from the soil is difficult because the contrast between soil particles and roots often presents challenges to segmenting for root extraction. In this paper, we proposed a fully automated method based on convolutional neural networks, called SegRoot, adapted for segmenting root from complex soil background. Our method eliminates the need for delicate feature designing which requires significant expert knowledge. The trained SegRoot networks learned morphological features with different abstraction levels directly from root images. Thus, the generalization and adaptation of the proposed method was expected across different root images. Using images of soybean roots, high performance of segmentation results were obtained by our benchmark SegRoot with testing dice score of 0.6441 (where 1 is a perfect score). When compared with human traced root lengths, an excellent correlation in root total length estimation was achieved with R2 of 0.9791. We also applied SegRoot to images with entirely different soil type collected from a forest ecosystem. Even without training on those images, a good generalization capability was obtained. Additionally, the impact of network capacity was also studied in order to find a cost-effective network suitable for in field application. We believe this automated segmentation method will revolutionize the measurement of plant roots in soil by dramatically increasing the ability to extract data from minirhizotron images.