Applying colour-based feature extraction and transfer learning to develop a high throughput inference system for potato (Solanum tuberosum L.) stems with images from unmanned aerial vehicles after canopy consolidation

Mhango, J.K., Grove, I.G., Hartley, W., Harris, E. and Monaghan, J.M. (2021) Applying colour-based feature extraction and transfer learning to develop a high throughput inference system for potato (Solanum tuberosum L.) stems with images from unmanned aerial vehicles after canopy consolidation. Precision Agriculture.

[img] Text
Jim Monaghan Applying colour UPLOAD.pdf - Accepted Version
Restricted to Repository staff only until 18 September 2022.

Download (543kB) | Request a copy

Abstract

Potato (Solanum tuberosum) stem density variation in the field can be used to inform harvest timing to improve tuber size distribution. Current methods for quantifying stem density are manual with low throughput. This study examined the use of Unmanned Aerial Vehicle imagery as a high-throughput alternative. A colour-based feature extraction technique and a deep convolutional neural network (CNN) were compared for their effectiveness in enumerating apical meristems as a proxy to subtending stems. Two novel colour indices, named the cumulative blue differences index and blue difference normalized index, showed significant differences (P < 0.001) between meristematic leaves and mature leaves in comparison to other indices. The two indices were used to generate 500 pseudo-labelled human-corrected images as training data for the CNN. Benchmarked against a human labelled test dataset, the CNN performed better with a normalized Root Mean Square Error (nRMSE) of 0.09 than the sole use of the image analysis algorithm (nRMSE = 0.3) in predicting the number of meristems in a canopy at 52 days after planting. Furthermore, the CNN had better precision (Intersection over Union [IOU]: 0.49 and 0.56, respectively) than the image analysis algorithm (IOU: 0.33 and 0.13, respectively). Meristem counts in both approaches showed a linear relationship with actual subtending stem counts (P < 0.001). This study demonstrates the validity of using traditional image analysis and CNNs to generate meristem detectors with acceptable nRMSE. Transfer learning with CNN is proposed for developing meristem detectors for evaluating stem density variation from UAV images in the field.

Item Type: Article
Keywords: Vegetation indices, Deep learning, Potato, Plant population, Phenotyping, Machine vision
Divisions: Agriculture and Environment (from 1.08.20)
Depositing User: Mrs Rachael Harper
Date Deposited: 04 Oct 2021 10:16
Last Modified: 04 Oct 2021 10:16
URI: https://hau.repository.guildhe.ac.uk/id/eprint/17749

Actions (login required)

Edit Item Edit Item