Fine Root Image Processing Based on Deep Learning and Prior Knowledge

Abstract

In situ measurement of root traits is a very essential step for a better understanding of root morphological development, nutrition supply, and physiological process for precision agriculture. However, difficulty in acquisition of root traits and complicated constitutes of root hairs, root axes and low contrast background presents challenges for root traits segmentation. In this paper, an efficient and accurate model was developed for segmenting complicated root and root hair image by constructing region of interest and adding prior knowledge in convolutional neural network. Since roots appear on the image randomly and irregularly, regional growth result was used to position the root area and thus construct the region of interest. Transfer learning was applied to pre-trained model-weights on relevant dataset as initial parameters. Root axes and root hairs were separated using pruning method, and root parameters were extracted based on it. The result showed that the P-T-U-Net model (U-Net base on prior knowledge and transfer learning) had the best performance in root segmentation. The IOU, PA, and F1 was 0.881, 0.982, and 0.931, respectively. Segmentation for fine roots were a little worse than those for root axis, since root hairs are dramatically smaller than root axes. But averagely, it had F1 score over 0.9. Plant species has little influence on the segmentation. It was able to figure out details on crossing and overlapping roots, and was able to extract micro root changes over time.

Keywords: Root hair, Image segmentation, Transfer learning, Prior knowledge, Deep Learning