Can soil spectroscopy contribute to soil organic carbon monitoring on agricultural soils?
Abstract. Carbon sequestration in soils under agricultural use can contribute to climate change mitigation. However, the spatial-temporal monitoring of soil organic carbon (SOC) requires more efficient data acquisition. The use of soil Vis-NIR spectroscopy is a promising research field in this context. However, the interpretation of the recorded spectral signal with regards to SOC is not trivial due to the complexity of the soil matrix, and factors affecting the measurements under field conditions. A model-building process is required to relate the spectral signal to the SOC content. For this study, spectral on-the-go proximal measurements and soil sampling were conducted on a long-term field experiment (LTE) located in the state of Saxony-Anhalt, Germany. SOC values ranged between 14–25 g kg−1 due to different fertilization treatments. Partial least squares regression (PLSR) models were built on behalf of spectral laboratory and field measurements conducted with two spectrometers and preprocessed by various methods. A data correction of the field data was done with three different approaches: linear transformation, piecewise direct standardization (PDS), and external parameter orthogonalization (EPO). The models were then thoroughly interpreted with regards to spectral wavelength importance using regression coefficients (RC) and variable importance in projection scores (VIP). The detailed wavelength importance analysis disclosed the challenge of using soil spectroscopy for SOC monitoring. The use of spectrometers with a differing spectral resolution for soil Vis-NIR measurements under varying soil conditions revealed shifts in wavelength importance. Still, some wavelengths related to SOC were extracted (560 nm, 1330 nm, 1400 nm, 1720 nm, and 1900 nm) by various preprocessing methods and were highly important in models trained on both, laboratory, and field measurements. Furthermore, we showed, that the correction of spectral field data with spectral laboratory measurements improved the predictive performance of the models built on behalf of the proximal on-the-go sensing measurements.
Javier Reyes and Mareike Ließ
Status: final response (author comments only)
RC1: 'Comment on egusphere-2022-273', Anonymous Referee #1, 10 Jul 2022
- AC1: 'Reply on RC1', Javier Reyes, 11 Nov 2022
RC2: 'Comment on egusphere-2022-273', Anonymous Referee #2, 02 Oct 2022
- AC2: 'Reply on RC2', Javier Reyes, 11 Nov 2022
Javier Reyes and Mareike Ließ
Javier Reyes and Mareike Ließ
Viewed (geographical distribution)
This paper aims to evaluate the ability to use soil vis-NIR spectroscopy to monitor SOC at the field scale. Two instruments (ASD and Veris) are used to acquire spectral data under laboratory and field conditions. Different spectral processing methods are combined with PLS regression to derive the best prediction models. In addition, spectral field data are corrected with the laboratory-based counterparts, which improves the performance of models built with field measurements. The topic is interesting and fits within the scope of SOIL. However, I think the paper needs a major restructuring and more clarity in the objectives and results section.
From the title, abstract (L5-10), and the introduction section (L75-80), I can see that this paper aims to evaluate the possibility of using vis-NIR to detect the spatial-temporal changes in SOC at the field scale. This is an interesting research question, because spatially continuous SOC monitoring in an efficient manner is important for studying SOC sequestration and formulating climate mitigation strategies. However, as I advanced in reading through the results and discussion sections, it seems to me that the current analysis actually focuses on ‘assessing the ability to use vis-NIR to predict SOC’’. Many studies on a field, regional or continental scale actually have done this, using PLS (current study), Memory-based machine learning, or Cubist models combing with different spectral processing techniques. Based on this, it is not clear to me what the main objective of this paper is, and thus the novelty of this paper.
If the authors focus on ‘spatially continuous SOC monitoring based on the long-term field experiment’, I think readers may want to know whether we can accurately detect changes in SOC using field vis-NIR spectroscopy (Veris), and what are the main factors that affect the accuracy of predictions. I am assuming that if the changes in SOC are small (e.g., based on a short-term fertilization experiment), then vis-NIR may be unable to capture it because of the higher prediction error based on the field spectral measurements (Figure 4 Veris). Furthermore, the addition of chemical fertilizers (L100) may affect the soil mineral matrix, e.g., by releasing the Fe minerals, which will strongly affect the spectral absorption features in the electronic transition region (425, 480, 513, 650, 903, 1000 nm). So, the treatments (fertilizer addition, and crop rotation… as described in L95) may lead to different interpretations in terms of model performance parameters, RC, and VIP. In my opinion, authors should focus on this perspective and pay more attention to what vis-NIR can do in detecting changes in SOC at spatial and temporal scales. If authors were to add new analysis in this regard, you may delete some previous analysis, i.e., comparing the model performance of 4 spectral processing methods (just pick the best one), because 1) multiple studies have shown that the effect of pre-processing on the accuracy of SOC predictions is very limited (Baldock et al., 2013; Dotto et al., 2018; Igne et al., 2010), and 2) too much information may distract from the key message of this paper.
In a summary, the manuscript is not appropriate for publication in SOIL in its present form. I suggest authors carefully consider the objectives of this paper and then reformulate the title, results, and conclusions.
Some minor comments:
Abstract Only 3 sentences are about the results and conclusions while concentrating too much on the description of the context and methods. Need to revise.
L35-42 in the introduction, the authors described the continuous SOC monitoring experiments in detail, so I think this is relevant to your research objectives. but…(see general comment).
L60 be more specific, preferably by providing examples.
L158 I suggest adding the ‘Ratio of Performance to Deviation (RPD)’ as an additional model performance parameter. This is a robust, widely used indicator to judge spectral models. (Change et al., 2001; Viscarra Rossel et a., 2012 EJSS).
L195-197 R2 < 0.5, PRD < 1.4? it seems an un-reliable model, please check with RPD values.
L330 revise this conclusion after 1) calculating RPD (previous comment) to judge model performance; 2) adding analysis regarding “monitoring change” (see general comment)
L250-270 consider that add statistical analysis to compare the differences in model performance between different spectral processing methods and spectral correction methods (L306)
L330, L345 be careful with these conclusions: given the low model performance of Veris field IR (Fig. 4 Veris field-SGCR, R2< 0.5), small changes in SOC may be undetectable using vis-NIR.
Figure cation: explain the elements of boxplot, e.g., min, max, Q1 Q3 outliers...
Figure 3, 5: consider combining the plots that belong to the same categories, e.g., A with B, C with D (using colors to distinguish them).
Figure 9, 10: round the wavelength numbers
Figures 7, 8, 9, 10. If I understand correctly, Figure 9 (median values) only carried part of the information that was already given by Figure 7 (median + range)? If yes, why not combine them? Also for Figure 8 and Figure 10.