The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data

Ng, Wartini; Minasny, Budiman; Mendes, Wanderson de Sousa; Demattê, José Alexandre Melo

doi:https://doi.org/10.5194/soil-6-565-2020

Articles | Volume 6, issue 2

https://doi.org/10.5194/soil-6-565-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/soil-6-565-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 6, issue 2

Original research article

|

17 Nov 2020

Original research article |

| 17 Nov 2020

The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data

Wartini Ng, Budiman Minasny, Wanderson de Sousa Mendes, and José Alexandre Melo Demattê

Download

Final revised paper (published on 17 Nov 2020)
Preprint (discussion started on 17 Sep 2019)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review: Estimation of effective calibration sample size using visible near infrared spectroscopy: deep learning vs machine learning by Wartini Ng et al', Anonymous Referee #1, 19 Oct 2019
- AC1: 'Response', Wartini Ng, 01 Nov 2019
RC2: 'additional comment to review: Estimation of effective calibration sample size using visible near infrared spectroscopy: deep learning vs machine learning by Wartini Ng et al', Anonymous Referee #1, 19 Oct 2019
- AC3: 'Response', Wartini Ng, 08 Apr 2020
RC3: 'Referee Comment', Anonymous Referee #2, 16 Mar 2020
- AC2: 'Response', Wartini Ng, 08 Apr 2020

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

ED: Reconsider after major revisions (14 Apr 2020) by Bas van Wesemael

AR by Wartini Ng on behalf of the Authors (12 May 2020) Author's response Manuscript

ED: Reconsider after major revisions (20 May 2020) by Bas van Wesemael

ED: Referee Nomination & Report Request started (02 Jun 2020) by Bas van Wesemael

RR by Anonymous Referee #2 (02 Jul 2020)

RR by Anonymous Referee #3 (27 Aug 2020)

Suggestions for revision or reasons for rejection

The paper discusses the performance of convolutional neural networks (CNN) compared to traditional machine learning techniques in function of the number of calibration samples. The first reviewer of the previous version misunderstood the objectives and therefore had serious doubts on the novelty and the strategies. The authors have changed the title and clarified the objectives in order to deal withthis misunderstanding.
The second reviewer mentions a lack of a discussion section. This has now been added.
Specific comments
Line 59, 85-86, 130, 139 …. (please check throughout the anuscript) …spectral data… or ‘spectra’ on its own.
Line 59-60 …predict clay content..
Line 64 …soil spectral libraries…
Line 68 What are ‘increased calibration samples’? Is there a word missing?
Line 68 Split up the sentence: …performance. However, there ….
Line 70 A strategy …
Line 73 …. How many samples….
Line 75 CNN will outperform traditional…
Line 78 …CNN model to outperform machine….
Lines 81, 82 Delete either ‘specifically’ or ‘specific’
Line 83 ..models will reach..
Line 87 There seems to be a word missing : …achieved wen the number…
Line 92 Delete ‘of’
Line 95 …and sedimentary rocks..
Line 95 Delete ‘samples’
Line 102 In general it is better to work as much as possible with the original data. There is no need to multiply all organic carbon contents with 1.724. I would prefer if you use soil organic carbon (SOC) for all prediction models. After the Van Bemmelen factor is an empirical one and adds unnecessary noise to the data.
Line 103 .. to extract exchangeable aluminium, calcium and magnesium…
Line 117 …spectral measurement…
Line 152 …as one-dimensional data…
Line 156 Word missing : …was trained using a batch…
Line 200 ….each filter of…
Line 260 I do not understand what you want to illustrate in Fig. 7. The sensitivity analysis is shown in Fig.8.
Line 333 Delete ‘in this paper’

Hide

ED: Publish subject to minor revisions (review by editor) (27 Aug 2020) by Bas van Wesemael

AR by Wartini Ng on behalf of the Authors (28 Aug 2020) Author's response Manuscript

ED: Publish subject to technical corrections (14 Sep 2020) by Bas van Wesemael

ED: Publish subject to technical corrections (30 Sep 2020) by Kristof Van Oost (Executive editor)

AR by Wartini Ng on behalf of the Authors (01 Oct 2020) Manuscript

Short summary

The number of samples utilised to create predictive models affected model performance. This research compares the number of samples needed by a deep learning model to outperform the traditional machine learning models using visible near-infrared spectroscopy data for soil properties predictions. The deep learning model was found to outperform machine learning models when the sample size was above 2000.