Can Basic Soil Quality Indicators and Topography Explain the Spatial Variability in Agricultural Fields Observed from Drone Orthomosaics?

Round 1

Reviewer 1 Report

This study evaluated the relationship between the spatial variability of agricultural fields observed from drone orthomosaics and soil quality indicators. The topic is interesting, however, some important problems were not considered in their methods. Therefore, I could not recommend it publish on the Agronomy in its current state. The detail comments are as follows.

1.     The orthomosaics of remote sensing data in the agricultural fields are related to many factors such as vegetation type, planting time, field spacing, fertilization, pests and diseases of plants, and so on, which are far more than those two factors mentioned in your manuscript.

2.     What is the purpose of interpolation? In addition, the IDW method is not suitable for every soil parameters. The image obtained from drone has high spatial resolution at centimeter level, which can match the field observations well.

3.     Section 3.1   The significance of correlation coefficients should be tested.

Author Response

Answers:

Thank you for your comments. Below answers your detailed questions. Furthermore, English editing has been done by a native speaker based on your request.

1. It is true that remote sensing data from agricultural fields are related to many mentioned factors. However, the objective of our study was to evaluate the relationship between multiannual drone orthomosaic datasets and commonly available soil and field indicators and evaluate how well they can explain spatial variability of agricultural fields. Fields were managed evenly during the experiments meaning that vegetation type or planting time could not cause within-field variability to field. This information has been now emphasized in Material section (2.1) by adding there the following sentence: “All fields were managed evenly during the experiments, i.e. farming practices such as sowing and fertilizing were not expected to cause within-field variability to the fields.” Field spacing causes spatial variability in high resolution, thus we resampled our data to 5 m GSD, to minimize this effect . Furthermore, we stated in our discussion that “The reason for crop related variation in drone images can also be caused by other aspects than soil, such as lodging, mistakes in crop sowing and management, or extreme weather conditions (wetness/dryness)”. As you mentioned pests and diseases, we have inserted them into the above-mentioned sentence.
2. We selected the IDW as it provided the best suitable interpolation results for our low sample size. We added in 2.2.5 more details:

“We selected Inverse Distance Weighting (IDW) spatial interpolation method following the guidelines given by Li and Heap (2014). We also considered the geostatistical method ordinary kriging and, based on variograms, the soil samples followed spatial autocorrelation but due to the small number of samples this structure was not always so clear. We used the classical IDW (Shepard 1968) with the distance coefficient P = 2, which is the most used version of IDW (Bărbulescu et al. 2021) and which visually provided suitable results. Radočaj et al. (2021) also observed similar interpolation accuracy for soil parameters using ordinary kriging and IDW.”

Our aim was to be able to study the within-field variability in the entire field. As drone orthomosaics and topographical wetness index were already available for the entire field parcel, we decided to apply it also to soil data. The soil data is not expected to present high spatial variation around centimeter distance. Thus, we interpolated the field samples to a 5 meters grid. The interpolated grid was created to increase the number of training data for the regression analysis. Besides, the resample of the drone data was also done to remove possible small-scale effects caused, for instance, by tractor lines.

Furthermore, we tested another popular spatial interpolation method, ordinary kriging (OK), and analyzed if it could improve the results. At first, we carried out a grid search to find optimal parameters (variogram_model: ["linear", "power"], nlags: [4, 6, 8, 15, 20, 25]) for ordinary kriging. If spatial autocorrelation was not found, the kriging was not possible to calculate (19 cases out of 52). We calculated the correlations between all orthomosaics and kriging interpolated rasters and compared them to IDW (Figure 5 in manuscript). The differences between the correlations were generally small (less than 0.1). The highest difference was found between variable pen_05-15 and dataset 2020-07-16_ms_i3 (0.56). As an average, the higher absolute correlations were found using IDW for field l2 and OK for field i3. For fields l1 and i4, the average absolute correlations were similar. Therefore, it was concluded that OK interpolation strategy did not yield better results than IDW.

3. The significance of correlation results based on Wald Test with t-distribution (p<0.05) information is now added to figure 5.

Bărbulescu, A., Șerban, C., & Indrecan, M. L. (2021). Computing the beta parameter in IDW interpolation by using a genetic algorithm. Water, 13(6), 863.

Li, J., & Heap, A. D. (2014). Spatial interpolation methods applied in the environmental sciences: A review. Environmental Modelling & Software, 53, 173-189.

Radočaj, D., Jug, I., Vukadinović, V., Jurišić, M., & Gašparović, M. (2021). The Effect of soil sampling density and spatial autocorrelation on interpolation accuracy of chemical soil properties in arable cropland. Agronomy, 11(12), 2430.

Shepard, D. (1968). A two-dimensional interpolation function for irregularly-spaced data. In Proceedings of the 1968 23rd ACM national conference (pp. 517-524).

Author Response File: Author Response.docx

Reviewer 2 Report

The paper "Can basic soil quality indicators and topography explain the spatial variability of agricultural fields observed from drone orthomosaics?" presents a comprehensive analysis of how soil characteristics can be used to interpret drone recordings in agriculture in Finland environmental conditions. The authors provide a thorough review of the state of the art in the field and present the method for extracting soil information from drone recordings. The method is based on a combination of machine learning techniques and statistical analysis, which is applied to a dataset of soil samples and corresponding drone recordings. The results show that the proposed method is able to accurately predict soil characteristics from drone recordings, demonstrating the potential of this approach for soil mapping and precision agriculture.

The paper is well written and well organized, making it easy to follow the authors' arguments and the methods used. The methodology is clearly explained, and the results are presented in a clear and concise manner. The authors also discuss the limitations of the method and suggest future directions for research.

Overall, this paper provides a valuable contribution to the field of precision agriculture and is a useful resource for researchers and practitioners interested in using drone recordings for soil mapping and analysis. The authors' approach is innovative and has the potential to be widely adopted in the industry.

But there are a few questions remaining.

1.       Why you resample soil data by spatial interpolation method?

2.       When you did interpolation, which included a prediction error depending on the spatial density of the acquired data, soil heterogeneity, tuned parameters, etc, and subsequently compared the interpolated data (which is basically not true values) with the drone data, you did ML regression on the hypothesized data. In my opinion, this is a significant methodological fault. It would be better to include several interpolation techniques and compare regression results that way.

3.       What was the criterion for IDW selection and p=2?

4.       Why you use deterministic instead geostatistical method of spatial data interpolation?

Author Response

Thank you for your comments. We have carefully considered them and have revised the section 2.2.5 related to interpolation. We also performed the experiment using geostatistical ordinary kriging method and checked if it has an effect to correlations. Furthermore, English editing has been done by a native speaker based on the request of another reviewer. Answers for your detailed questions:

1. Our aim was to be able to study the within-field variability in the entire field. As drone orthomosaics and topographical wetness index were already available for the entire field parcel, we decided to apply it also to soil data. The interpolated grid was also increasing the number of training data for the regression analysis (See also answer 4).
2. It is true that spatial interpolation includes a prediction error. In section 4.3, we already discussed the limitations related to soil data interpolation. We emphasized it more clearly by adding the following sentence there: “It should be noted that interpolated soil data includes uncertainties, which can impact the results.” We selected the soil sampling locations based on spatial variability identified in orthomosaics (as mentioned in section 2.1.3) and we believe that we managed to handle the soil heterogeneity well at least on a large scale. Spatial data interpolation is a way (not optimal) to produce training data for machine learning and it has been recently applied for example by Fu et al. (2023), Pradhan et al. 2022 and Nevavuori et al. (2022). We also performed the experiment using geostatistical ordinary kriging method and checked if it has an effect on correlations (See below Kriging test, Figure 1).
3. We used the classical IDW (Shepard 1968) with the distance coefficient P = 2 and which is the most used version of the IDW (Bărbulescu et al. 2021) and which resulted in a visually comprehensive result. This information has been added also to the manuscript.
4. The number of samples available was small and Inverse Distance Weighting (IDW) provided best visual results to all parameters. We tested ordinary kriging but it did not work well on all parameters and it gave similar correlations for the parameters that the interpolation worked. Please see the details below.

We added now details in 2.2.5:

“We selected spatial interpolation method Inverse Distance Weighting (IDW) following the guidelines given by Li and Heap (2014). We also considered the geostatistical method ordinary kriging and, based on variograms, the soil samples followed spatial autocorrelation but due to the small number of samples this structure was not always so clear. We used the classical IDW (Shepard 1968) with the distance coefficient P = 2, which is the most used version of IDW (Bărbulescu et al. 2021) and which visually provided suitable results. Radočaj et al. (2021) also observed similar interpolation accuracy for soil parameters using ordinary kriging and IDW.”

Kriging test

Furthermore, we tested another popular spatial interpolation method, ordinary kriging (OK), and analyzed if it could improve the results. At first, we carried out a grid search to find optimal parameters (variogram_model: ["linear", "power"], nlags: [4, 6, 8, 15, 20, 25]) for ordinary kriging. If spatial autocorrelation was not found, the kriging was not possible to calculate (19 cases out of 52). We calculated the correlations between all orthomosaics and kriging interpolated rasters and compared them to IDW (Figure 5 in manuscript). The difference between correlation were generally small (less than 0.1). The biggest difference was found between variable pen_05-15 and dataset 2020-07-16_ms_i3 (0.56). As an average, the higher absolute correlations were found using IDW for field l2 and OK for field i3. For fields l1 and i4, the average absolute correlations were similar. Therefore, it was concluded that OK interpolation strategy was not yielding better results.

Figure 1. The correlation coefficient between drone datasets and soil properties (interpolation algorithm: Ordinary Krigin) for all datasets collected in experimental fields l1, l2, i3 and i4.  A= topsoil biological indicators, B= subsoil macroporosity, C= compacted layers in soil profile, D= topsoil structure, E= subsoil structure, pen05−15 = mean soil penetration resistance in the layer of 5−15 cm, and pen20−40 = in the layer of 20−40 cm, SOC = soil organic carbon content, C_SOC = clay/soil organic carbon ratio in 0−20 cm layer.

References

Bărbulescu, A., Șerban, C., & Indrecan, M. L. (2021). Computing the beta parameter in IDW interpolation by using a genetic algorithm. Water, 13(6), 863.

Fu, Y., Cheng, Q., Jing, L., Ye, B., & Fu, H. (2023). Mineral Prospectivity Mapping of Porphyry Copper Deposits Based on Remote Sensing Imagery and Geochemical Data in the Duolong Ore District, Tibet. Remote Sensing, 15(2), 439.

Li, J., & Heap, A. D. (2014). Spatial interpolation methods applied in the environmental sciences: A review. Environmental Modelling & Software, 53, 173-189.

Nevavuori, P., Narra, N., Linna, P., & Lipping, T. (2022). Assessment of Crop Yield Prediction Capabilities of CNN Using Multisource Data. In New Developments and Environmental Applications of Drones: Proceedings of FinDrones 2020 (pp. 173-186). Springer International Publishing.

Pradhan, B., Jena, R., Talukdar, D., Mohanty, M., Sahu, B. K., Raul, A. K., & Abdul Maulud, K. N. (2022). A New Method to Evaluate Gold Mineralisation-Potential Mapping Using Deep Learning and an Explainable Artificial Intelligence (XAI) Model. Remote Sensing, 14(18), 4486.

Radočaj, D., Jug, I., Vukadinović, V., Jurišić, M., & Gašparović, M. (2021). The Effect of soil sampling density and spatial autocorrelation on interpolation accuracy of chemical soil properties in arable cropland. Agronomy, 11(12), 2430.

Shepard, D. (1968). A two-dimensional interpolation function for irregularly-spaced data. In Proceedings of the 1968 23rd ACM national conference (pp. 517-524).

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have revised the manuscript carefully.