Research output per year
Research output per year
Yuxin Ma*, Budiman Minasny, Alex McBratney, Laura Poggio, Mario Fajardo
Research output: Contribution to journal › Comment/Letter to the editor › Academic
We would like to draw your attention that we had misreported a citation in our paper (Ma et al., 2021). This miscitation does not affect the results and conclusions. Nevertheless, we would also add a discussion to the paper based on feedback we received. In section 2.4.2. “Depth intervals converted to point depth – 3 D” of Ma et al. (2021), we wrote one of the methods to convert from depth interval observations to point observations is: 3) Fine discretization: Depth values were assigned at every centimeter of the depth intervals (Filippi et al., 2019) (Fig. 2c). The above fine discretization method was not originated by Filippi et al. (2019), and we apologise for that assertion. That reference should be deleted. Filippi et al. (2019) first splined the soil pH data at every 1 cm and then fit the 3D model using the splined values. To clarify it, we calculated predictions based on 1 cm splined SOC (%) as done by Filippi et al. (2019). We compared the results with one of the approaches in assigning point-depth values mentioned in our paper, i.e., method 3D-B (average depth + boundaries, Fig. 2b). The Quantile Regression Forest (QRF) model was fitted to the data. The predictions of SOC with depth based on splined values were smoother than the ‘average depth + boundaries’ assignment but still show a jagged trend, e.g., soil profile: ed187 (Fig. 1). Thus, generally, a tree-based model (e.g., random forest) will create a stepped artefact as it is not a smooth function. Including more fine depth increment (i.e., every 1 cm) will create a smoother depth prediction and thus the approach of Filippi et al. (2019) reduces the artefact. However, we should also caution the use of fine depth increment as inputs to any model as they were interpolated. An observation with larger depth interval (e.g., 60–100 cm) would be weighted more in the model compared to smaller depth interval observations (e.g., 0–5 cm). As the influence of depth is very large in the model, using excessive point-depth will force the model to overfit the depth function. Another approach of “data augmentation” was suggested by (Roudier et al., 2020), where a random value of observations (depth and soil observation) was drawn from the observed depth intervals. Still, one had to recognise the uncertainty of the augmented data from such an approach. We reiterate that future efforts should be focused on more efficiently collecting depth soil data rather than circumventing it with machine learning models. The publisher would like to apologise for any inconvenience caused.
Original language | English |
---|---|
Article number | 115631 |
Journal | Geoderma |
Early online date | Dec 2021 |
DOIs |
|
Publication status | Published - 15 Mar 2022 |
Research output: Contribution to journal › Article › Academic › peer-review