A Dynamic Hierarchical Bayesian Approach for Forecasting Vegetation Condition
- 1The Data Intensive Science Centre, Department of Physics and Astronomy, University of Sussex, Brighton BN1 9QH, UK
- 2Astronomy Centre, Department of Physics and Astronomy, University of Sussex, Brighton BN1 9QH, UK
- 3Department of Geography, School of Global Studies, University of Sussex, Brighton BN1 9QJ, UK
- 1The Data Intensive Science Centre, Department of Physics and Astronomy, University of Sussex, Brighton BN1 9QH, UK
- 2Astronomy Centre, Department of Physics and Astronomy, University of Sussex, Brighton BN1 9QH, UK
- 3Department of Geography, School of Global Studies, University of Sussex, Brighton BN1 9QJ, UK
Abstract. Agricultural drought, which occurs due to a significant reduction in the moisture required for vegetation growth, is the most complex amongst all drought categories. The onset of agriculture drought is slow and can occur over vast areas with varying spatial effects, differing in areas with a particular vegetation land cover or specific agro-ecological sub-regions. These spatial variations imply that monitoring and forecasting agricultural drought require complex models that consider the spatial variations in a given region of interest. Hierarchical Bayesian Models are suited for modelling such complex systems. Using partially pooled data with sub-groups that characterise spatial differences, these models can capture the sub-group variation while allowing flexibility and information sharing between these sub-groups. This paper's objective was to improve the accuracy and precision of agricultural drought forecast in spatially diverse regions with a Hierarchical Bayesian Model. Results showed that the Hierarchical Bayesian Model was better at capturing the variability for the different agro-ecological zones and vegetation land covers compared to a regular Bayesian Auto-Regression Distributed Lags model. The forecasted vegetation condition and associated drought probabilities were more accurate and precise with the Hierarchical Bayesian Model at 4 to 10 weeks lead times. Forecasts from the hierarchical model exhibited higher hit rates with a low probability of false alarms for drought events in semi-arid and arid zones. The Hierarchical Bayesian Model also showed good transferable forecast skills over counties not included in the training data.
Edward E. Salakpi et al.
Status: final response (author comments only)
-
RC1: 'Comment on nhess-2021-290', Anonymous Referee #1, 22 Dec 2021
Review of: A Dynamic Hierarchical Bayesian Approach for Forecasting Vegetation Condition. This study proposed a method to forecast vegetation condition index by means of soil moisture and precipitation data in Kenya by means of Bayesian method. I find the study not very well structured and difficult to follow given lack of information on the data sources used and period of analysis. In any case the justification of the study is mostly related to the assessment of drought forecasting but I cannot find any analysis and result that can be directly focused on drought analysis. The authors should detail better and illustrate more if long-time series are used and if forecasting is useful during critical drought periods. See specific comments below. They refer to specific lines of the manuscript:
- Does this number refer to the global scale? What is the source?
- Note that enhanced atmospheric evaporative demand increases the severity of agricultural droughts, particularly under low soil moisture conditions.
- But also human practices may reinforce drought severity or reduce it as a function of soil management, crop rotation. See the several studies by Prof. Rattan Lal about this issue.
- I would say that ecological droughts and hydrological droughts are also strongly complex. I would qualify this issue.
24-24. I would include issues related to crop practices, crop types, etc.
36-88. In the introduction I am surprised that dynamic forecasting of drought based on climate models is not considered (e.g. https://www.nature.com/articles/s41612-021-00189-4). There is a large research topic on this issue that should be mentioned/discussed in relation to the statistical techniques proposed in this article.
- What are the sources of precipitation, VCI and soil moisture? This must be detailed and discussed in relation to the availability of the data, quality, time period of analysis, etc. There is not sufficient information to determine which data the authors are using.
- What are the variables that are intended to be forecasted (precipitation? Soil moisture? VCI? All of them?) This is not clear in the methodology what is the target variable and what is the usefulness of the other variables. Maybe are precipitation and soil moisture possible drivers of the VCI and they are used to generate a predictive model of VCI? This should be clarified. This is indicated below 180-220; but it should be explained in detail before to avoid confusion. I would also suggest atmospheric evaporative demand as predictor of the VCI as several studies have demonstrated strong importance on agricultural and ecological drought conditions.
Figure 5. Why is only one year of results showed in this plot? I would suggest to include longer time series as we cannot determine if this year correspond to drought or to normal and humid conditions.
Figure 6. What about seasonality? Is the model capable to forecast vegetation conditions with the same accuracy for seasons of high or low vegetation activity? This should be at least discussed. What is the different performance between drought and non-drought years?
Figure 7. I see not only agricultural areas are considered in the study but also forest lands so introduction should be modified to also focus on ecological droughts.
Basically all the figures are showing a single year for the VCI forecasting and it is not clear for me if this is the period of analysis or it is an arbitrary selection. In any case. This must be clarified as I do not think robust results and conclusions related to the suggested methodological approach can be obtained from the application to a single year.
- Where is this analysed? I cannot find a plot in which this relationship is showed.
- A reason to include also atmospheric evaporative demand as predictor...
-
AC1: 'Reply on RC1', Edward Salakpi, 08 Apr 2022
Reviewer 1
Review of: A Dynamic Hierarchical Bayesian Approach for Forecasting Vegetation Condition. This study proposed a method to forecast vegetation condition index by means of soil moisture and precipitation data in Kenya by means of Bayesian method. I find the study not very well structured and difficult to follow given lack of information on the data sources used and period of analysis. In any case the justification of the study is mostly related to the assessment of drought forecasting but I cannot find any analysis and result that can be directly focused on drought analysis. The authors should detail better and illustrate more if long-time series are used and if forecasting is useful during critical drought periods. See specific comments below. They refer to specific lines of the manuscript:
Response:
This paper is part 2 of a previous paper (https://doi.org/10.5194/nhess-2021-223) and uses the same data (reference was made to this in section 2.1.). But I realise this was the wrong decision, so a full description of the data used for this research will be added to the data section.
Secondly, the entire premise of this paper was based on our earlier paper, Barrett et al. 2020 (https://www.sciencedirect.com/science/article/pii/S003442572030256X). We aimed to improve the precision and accuracy of the VCI forecast beyond the six weeks achieved in the Barrett et al. 2020. Thus, the analysis of our results mainly focused on the accuracy and precision of the VCI forecast beyond six weeks. And more specifically, on how our proposed model differentiated VCI forecast within a given region of interest.
The research was in partnership with the NDMA. They are currently using VCI for monitoring drought. They have used the indicator extensively for their monthly drought reports and bulletins.
However, all the issues raised about the paper structure are well noted and will be addressed accordingly.
- Does this number refer to the global scale? What is the source?
Response:
Yes, this is on a global scale Source: https://wwfint.awsassets.panda.org/downloads/drought_risk__wwf_.pdf . The source of the information was not properly cited, and this will be fixed in the updated version of the paper.
- Note that enhanced atmospheric evaporative demand increases the severity of agricultural droughts, particularly under low soil moisture conditions.
Response:
We agree that atmospheric evaporative demand is vital for agricultural drought studies. Information on atmospheric evaporative demand will be included in the introduction and discussion and recommended for future work.
- But also human practices may reinforce drought severity or reduce it as a function of soil management, and crop rotation. See the several studies by Prof. Rattan Lal about this issue.
Response:
Comment accepted and well noted; information on this will also be included in the paper's literature review section.
- I would say that ecological droughts and hydrological droughts are also strongly complex. I would qualify this issue.
Response:
This is also well noted and will be included in the literature review section of the paper.
24-24. I would include issues related to crop practices, crop types, etc.
Response:
Comment well noted and accepted; issues on crop practices will be included in the introduction section of the paper.
36-88. In the introduction, I am surprised that dynamic forecasting of drought based on climate models is not considered (e.g. https://www.nature.com/articles/s41612-021-00189-4). There is a large research topic on this issue that should be mentioned/discussed in relation to the statistical techniques proposed in this article.
Response:
Comment well noted; however, the focus of this paper was using the Hierarchical Bayesian Model to provide differentiated forecasts within a given region of interest and I do not see how this links to the suggested paper. This notwithstanding, we will have a thorough look at the suggestions and see how to put them in context.
- What are the sources of precipitation, VCI and soil moisture? This must be detailed and discussed in relation to the availability of the data, quality, time period of analysis, etc. There is not sufficient information to determine which data the authors are using.
Response:
We accept that it was a wrong decision to refer the reader to data sources in a previous paper (https://doi.org/10.5194/nhess-2021-223). This will be corrected in the final draft of the paper. Please see the data sources below:
Data
Source (Producer)
Precipitation
Climate Hazards Group InfraRed
Precipitation (CHIRPS) - 2001-2018
Soil Moisture
European Space Agency’s
Climate Change Initiative (CCI) - 2001-2018
Surface Reflectances
NASA MODIS (MCD43A4 v006) - 2001-2018
- What are the variables that are intended to be forecasted (precipitation? Soil moisture? VCI? All of them?) This is not clear in the methodology what is the target variable and what is the usefulness of the other variables. Maybe are precipitation and soil moisture possible drivers of the VCI and they are used to generate a predictive model of VCI? This should be clarified. This is indicated below 180-220; but it should be explained in detail before to avoid confusion. I would also suggest atmospheric evaporative demand as predictor of the VCI as several studies have demonstrated strong importance on agricultural and ecological drought conditions.
Response:
The target variable is the VCI and lags of precipitation and soil moisture are being used as drivers for the VCI in this context. The comment is well noted and will be clarified as suggested.
In this paper we considered variables that could be directly derived from satellite remote sensing products. However, the suggestion of the atmospheric evaporative demand is well noted and will be recommended for future work
Figure 5. Why is only one year of results showed in this plot? I would suggest to include longer time series as we cannot determine if this year correspond to drought or to normal and humid conditions.
Response:
Unfortunately data was limited by access to readily available soil moisture data. At the time of data acquisition and preprocessing our soil moisture data was only up to 2018. HBM was calibrated to data from 2001 to 2016 and evaluated on 2017 and 2018 data which is why the model evaluation was a single year.
Figure 6. What about seasonality? Is the model capable to forecast vegetation conditions with the same accuracy for seasons of high or low vegetation activity? This should be at least discussed. What is the different performance between drought and non-drought years?
Response:
This comment is also well noted, analysis of seasonality is available and will be included in the discussion section of the paper.
Figure 7. I see not only agricultural areas are considered in the study but also forest lands so introduction should be modified to also focus on ecological droughts.
Basically all the figures are showing a single year for the VCI forecasting and it is not clear for me if this is the period of analysis or it is an arbitrary selection. In any case. This must be clarified as I do not think robust results and conclusions related to the suggested methodological approach can be obtained from the application to a single year.
Response:
This is also well noted, the definition of ecological drought will be included in the introduction section. We used a single year because the aim here was to show how well the HBM forecasted VCI for different agro-ecological zones and vegetation land covers for up to 12 weeks ahead.
- Where is this analysed? I cannot find a plot in which this relationship is showed.
Response:
These can be seen in the Percentage Relative Importance plots figures D1 and D2. These figures will be moved to the main document under the results section and discussed properly.
- A reason to include also atmospheric evaporative demand as predictor...
Response:
Well noted, It will be considered in future work as the funding phase for this project has ended.
-
RC2: 'Comment on nhess-2021-290', Anonymous Referee #2, 26 Feb 2022
General comments
The paper has objectives similar to those of the paper by Salakpi et al., NHESSD, “Forecasting vegetation conditions with a Bayesian auto-regressive distributed lags (BARDL) model (https://doi.org/10.5194/nhess-2021-223 ). The aim is to forecast vegetation conditions leading to agricultural droughts. The paper presents a new method, as compared to the previous paper, based on a dynamic Hierarchical Bayesian Model (HBM). The data used in the two papers are the same, and the method for assessing the forecasts are also similar. The HBM model has the advantage of providing differentiated forecasts within the same region, as compared to BARDL that provides homogeneous forecasts within each region. The BARDL model is used as benchmark for assessing the improvements provided by the HBM model. There are therefore lots of repetitions between the two papers. The present one refers to the previous one to present the data used in the study. This is not acceptable, as a paper should be self-contained and a minimum of information should be provided about data. The problem would be solved if the two papers were merged in a unique paper.
Concerning the present paper, I find it is well written and the analysis is sounded and honest. I only have a few minor comments listed below.
Specific comments
1/ P.1 line 16: the number of affected persons is different than in the previous paper
2/P.5. Section 2.1 should be developed
3/ p. 6 line 116: ROI is not defined
4/ p.7 line 135: the authors mention that seasonal means are subtracted, but how are the seasons defined?
5/ p.9 lines 172-175: this section is not clear
6/ Figure 4: I do not understand how to read this figure
7/ p.10 eq.(3). This is strange to present the equation of the model after the presentation of the HBM model that refers to the parameters listed in Eq. (3).
8/ p.12 line 253. The BARDL model based on a “no-pooling” is the same as the model presented in the previous paper, isn’t it? If this is true, this should be stated.
9/ p.12 line 261: the differences between the two models is Figure 5 are not so large. So the claim that HBM model is more accurate than BARDL is not fully justified.
10/ Revise Figures 6, 7, 9, 10, 13 as the full line and dotted lines cannot be distinguished.
11/ p. 20 section test transfer learning. From Figure 2, it is not easy to see which counties are used for calibration and for validation.
-
AC2: 'Reply on RC2', Edward Salakpi, 08 Apr 2022
Reviewer 2
General comments
The paper has objectives similar to those of the paper by Salakpi et al., NHESSD, “Forecasting vegetation conditions with a Bayesian auto-regressive distributed lags (BARDL) model (https://doi.org/10.5194/nhess-2021-223 ). The aim is to forecast vegetation conditions leading to agricultural droughts. The paper presents a new method, as compared to the previous paper, based on a dynamic Hierarchical Bayesian Model (HBM). The data used in the two papers are the same, and the method for assessing the forecasts are also similar. The HBM model has the advantage of providing differentiated forecasts within the same region, as compared to BARDL that provides homogeneous forecasts within each region. The BARDL model is used as benchmark for assessing the improvements provided by the HBM model. There are therefore lots of repetitions between the two papers. The present one refers to the previous one to present the data used in the study. This is not acceptable, as a paper should be self-contained and a minimum of information should be provided about data. The problem would be solved if the two papers were merged in a unique paper.
Concerning the present paper, I find it is well written and the analysis is sounded and honest. I only have a few minor comments listed below.
Response:
Thank you for the feedback, the issues on the omission of the data description are well noted and will be fixed in the final draft of the paper. The issue with repetitions is also noted and instead of merging, both papers will be published as parts A and B.
Specific comments
1/ P.1 line 16: the number of affected persons is different than in the previous paper
Response:
Comment well noted, this was an error and has been corrected in that paper, the correct figure is the one stated I this paper.
2/P.5. Section 2.1 should be developed
Response:
Comment noted and accepted, details on the data description will be added.
3/ p. 6 line 116: ROI is not defined
Response:
Comment well noted and will be fixed. This was an oversight and has been addressed
4/ p.7 line 135: the authors mention that seasonal means are subtracted, but how are the seasons defined?
Response:
That statement should read annual mean per land cover type and AEZ is subtracted. The error will be corrected
5/ p.9 lines 172-175: this section is not clear
Response:
Comment well-noted sentence should not have started with “However” that will be fixed and sentences in that section will be restructured to make it clear to the reader.
6/ Figure 4: I do not understand how to read this figure
Response:
The figure shows how the HBM is structured. The figure depicts how the model parameters and data are defined and how it all fits together. Maybe this is not too clear in the paper, the description of the figure will be elaborated in the caption of the figure.
7/ p.10 eq.(3). This is strange to present the equation of the model after the presentation of the HBM model that refers to the parameters listed in Eq. (3).
Response:
Comment well noted and this will be rearranged in a proper order that makes more sense
8/ p.12 line 253. The BARDL model based on a “no-pooling” is the same as the model presented in the previous paper, isn’t it? If this is true, this should be stated.
Response:
Yes it’s the same approach used in the previous paper, the comment is well-noted and will be well detailed in the paper.
9/ p.12 line 261: the differences between the two models is Figure 5 are not so large. So the claim that HBM model is more accurate than BARDL is not fully justified.
Response:
The statement will be toned down, rather the emphasis will be on the improvement in the lead times. Figure 5 will be dropped from the paper.
10/ Revise Figures 6, 7, 9, 10, 13 as the full line and dotted lines cannot be distinguished.
Response:
Comment well-noted figures will be replotted and made more legible to readers.
11/ p. 20 section test transfer learning. From Figure 2, it is not easy to see which counties are used for calibration and for validation.
Response:
Comment well-noted, the black boundary lines indicating the Non-Train counties in the figure will be made more visible. The maps will be redone.
-
AC2: 'Reply on RC2', Edward Salakpi, 08 Apr 2022
Edward E. Salakpi et al.
Edward E. Salakpi et al.
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
324 | 85 | 14 | 423 | 8 | 10 |
- HTML: 324
- PDF: 85
- XML: 14
- Total: 423
- BibTeX: 8
- EndNote: 10
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1