Mapping Flood-Based Farming Systems with Bayesian Networks

Round 1

Reviewer 1 Report

Review of: Mapping Flood-based Farming Systems with Bayesian networks   === Summary The authors present an analysis of remote sensed dataset with the goal of identifying flood-based farming system (FBFS) lands in different regions of eastern Africa. The data integration method used, a Bayesian network, attempts to take into account the uncertainties of the data and expert on-the-ground knowledge about which factors are most important in identifying suitability for FBFS.   === Major Comments The manuscript is a useful contribution to the broader literature on remote sensing-based land-use mapping as it demonstrates a new and successful application of methods being developed and applied elsewhere. The Bayesian network approach for incorporating disparate datasets is well-demonstrated here and the application to FBFS highlights an aspect of land use analysis that has not historically received much attention.   A few aspects of the manuscript should be revised to improve the overall impact and clarity of the work.   First, the distinction between hydrologic and agronomic flooding, as it relates to its representation in remote sensing data, should be clarified. It is not clear from the manuscript the importance or relevance of natural flooding events (overbank flow, ponding on the landscape, expansion of lake boundaries) in distinguishing between intentionally flooded areas and flooding that occurs by virtue of meteorological and/or hydrologic events beyond the effects of human intervention.   Similarly, the logical reasoning connecting the various factors that are proposed to be indicators FBFS should be made clearer and more explicit. In particular, the interactions of the various measures of moisture on the land surface with vegetation metrics make it difficult to determine what is important for the reader to take away. I would suggest a paragraph explaining the conceptualization (add it to section 2.2) as it relates to the data sets being used. For example, if it is believed that FBFS are commonly characterized by presence of flooding followed by increase in vegetation, longer duration dense vegetation cover, and low slopes, then explicitly explaining this prior to discussion of the Bayesian Network and data processing would greatly clarify the rest of the explanation.   Finally, if this work is being presented with the intent that it is useful for accurately mapping FBFS, some further discussion of accuracy is warranted. Specifically, the ability of the method to positively identify FBFS parcels along with it's ability to correctly identify non-FBFS parcels should be addressed. The discussion of uncertainty thresholds is a good foundation to extend from. For example, an evaluation of what uncertainty thresholds leads to most consistency with validation polygons would be a good indicator of the accuracy of the method/approach.   === Minor comments   p = page number l = line number (or range), otherwise referencing specific figure, table, section, or page   p2 l67: What is the nature of this diversity? Location on the landscape? The frequency of flooding? The types of crop? Some additional specificity here would help clarify.   p2 l69: On what basis are FBFS similar to other ecological systems? Perhaps the additional discussion of FBFS suitability or characteristics (mentioned in the Major Comments section above) would help elucidate this concept of ecological similarity as well.   p2 l76-77: "Agronomic flooding may also occur after the planting date..." Does agronomic flooding in FBFS occur prior to planting date (pre-irrigation, a means of wetting the soil profile in anticipation of planting, etc)? Is it more accurate to say that there is unpredictability in both the occurrence of both floods and how/if they are used for FBFS applications?   p3 l103: 567 km2 is not 4.5% of 2576 km2. Check for consistency here.   p3 Figure 1: It would be helpful to note some of the features described in lines 102-120  on these maps and images. Are there specific FBFS features to be seen in the selected images? If so, can these be noted too? Finally, Lake Victoria is referenced repeatedly in the text but is not shown on the maps - please add this and any other geographic points or landmarks of interest or relevance to the maps.   p4 Section 2.2: Add a discussion of the conceptualization of FBFS suitability factors or characteristics to this section, as mentioned in the Major Comments section (above)   p5, l156-158: Thank you for providing your code and workflow via GitHub links.   p5 l162: What is meant by reference to "emerging information" here and elsewhere in the text? It is not clear if there is an iterative process or how the information was "emerging".   p6 l182-184: What is meant by "variation in vegetation can either be due to flood or non-flood water"? When referring to "flood" here, is this meant to be agronomic flooding or meteorological flooding?   p6 l190-192: What value does the 'Suitable Soil' variable bring in terms of differentiating FBFS via the 'Flooded at Some Point' variable? In other words, why does 'Flooded at Some Point' depend on Suitable Soil instead of just the 'Water Presence' variable?   p6 l205: Explain how using 8-day composites affects the ability of the datasets to detect flooding at a temporal resolution sufficient to detect FBFS characteristics. It seems like the 8-day composites might introduce errors in the representation of transient features associated with flash floods or intermittent ponding.   p6-7: Sections 2.4-2.6: The citation of specific functions and software packages is appreciated, but this should be secondary to an explanation of the actual analysis performed. Perhaps the specific software citations could be summarized in a table in supplementary material to make room for a clearer description of the analysis in these sections.   p7 l239-240: What are the "emerging insights" mentioned here? It is unclear what is being referred to - please clarify.   p8 Section 2.6.1: How does filling depressions affect detection of relevant FBFS features? Might the DEM processing impact slope or flow accumulation distributions?   p8 Section 2.6.2: How is the vegetation sensitivity to water variation calculated? Please provide a quantitative explanation or formula if one was used.   p8 Section 2.6.3: It is not clear how the standard error of the NDFI relates to the sensitivity to flooding or what is meant by "surface states influence flooding via runoff generation". Is the concept that consistently wet soils consistently produce runoff while drier soils tend to exhibit more threshold-like runoff generation behaviors? Consider revising this paragraph for clarity.   p8 l309: "emerging raster" - What is meant by "emerging" here?   p12 Section 2.7.4: Explain what was being calculated or processed in addition to the function calls and software packages. For example, why was a "recursive partitioning algorithm in a decision tree classification" approach necessary?   p13 l418-433: How were FBFS identified in the validation data collection process, especially when differentiating between FBFS and conventional irrigation?   p13 l449: "distributions indicate that seasonality extends to water bodies and forest ecosystems" What is the basis for this assertion? Add text to clarify.   p15 l487-498: When referencing the location of certain results/phenomena in relation to regions or rivers, add notation for those regions or rivers to the maps for clarity.   p16 l510-514: Are floods, as described here, meant to refer to agronomic or hydrologic floods? Perhaps it would be good to provide a specific definition of terminology (flood, runoff, ponding, etc) at the beginning of the manuscript to help differentiate and clarify such references throughout.   p16 Figure 7: I'd suggest adding a line or other demarcation for Lake Victoria (and any other major water bodies) on these maps so that references to such are more easily identifiable.   p19 Section 3.4: Add a brief description of how the uncertainty metrics are calculated and why the Shannon entropy is the most appropriate measure.   p21 Table 4: The table indicates that conventional irrigated and rainfed fields represent a substantial portion of agricultural land. The FBFS potential is similarly high for these categories as it is for actual FBFS parcels. This warrants further discussion in the manuscript, specifically with an explanation supporting how the method is useful for positive and exclusive identification of FBFS (rather than being a more reliable identifier of agricultural land in general).   p22 Section 3.5.2: Add explanation or discussion of how accuracy is affected by the assumption of different uncertainty (optimistic vs pessimistic) prediction levels. How often does this method predict likely FBFS in cases where no FBFS are found? What conclusions can be drawn from these outcomes?   p23 l709-710: "we noticed that the experts were more comfortable in estimate the probabilities of extreme node states than for intermediate states..." What is the basis for this statement? How were expert opinions used beyond the initial delineation of variables? Add text to explain how/if the expert opinions were used for interpretation of results.   p23 l728: Missing reference

Author Response

Thank you for taking time to review our manuscript. Attached our response to your comments.

Author Response File: Author Response.docx

Reviewer 2 Report

I am satisfied with the topic and the approach to solving the problem. The conclusions are appropriate.

Author Response

Thank you for taking time to review our manuscript.

Reviewer 3 Report

The authors present a new approach for classifying land areas by their potential for flood-based farming systems (FBFS). The approach relies on a Bayesian network framework, which uses spatial data from satellite remote sensing to formulate Bayesian network nodes and categorize a location’s potential for FBFS based on a combination of spatial and temporal factors. The approach is illustrated on case studies of regions in Ethiopia and Kenya, and validated by comparing the modeled FBFS potential at a number of locations known to be true to FBFS or non-FBFS.

This work seems very relevant to the journal and may provide useful guidance for performing similar assessments of the spatial extents of specific irrigation or agricultural practices. The methodology seems sound, and the authors do a thorough job documenting their results and discussing the insights gained about the specific cases studied using their methodology. My primary concern with this manuscript is a lack of clarity in the description of the methodology. It seems that some important justifications and definitions are omitted, while a number of less important details (e.g., names of specific functions that were used in the data processing code) are included and may obscure the crux of the approach. In addition to some specific suggestions listed below, I recommend a thorough review of Sections 2-3 to ensure that the approach is described clearly enough to be replicable but without unnecessary detail, and that the use of all technical terms is consistent throughout the paper.

 

Comments on content:

Section 2.2, Figure 2 – The approach is not explained very clearly. For instance, the flow chart in the figure is not described in the text (e.g., acquisition of data, data discretization, etc. are not mentioned in the text). The difference between spatial and non-spatial components is not clear. Abbreviations used in Figure 2 are not defined in the caption. You should add a high-level description of the whole workflow shown in Figure 2 to this section. The output of the approach should also be clearer; my understanding that it aims to classify a pixel by “FBSF potential” – how exactly is that metric defined?

Section 2.3, Lines 160-162 – “We reviewed the essential literature related to the topic of FBFS and conducted high-level discussions with eleven experts to draft the Bayesian network, which we then updated based on farmer consultations and field observations.” I believe this statement deserves more discussion, to justify the structure of the Bayesian network presented by the authors. What literature was consulted? How did farmer consultations and field observations inform the structure of the Bayesian network? How were the specific nodes used in the network selected?

Section 2.3, Figure 3 – The text refers to “ten nodes” of spatial data, but the figure contains 17 total nodes, and the difference between spatial and non-spatial nodes isn’t clear from the figure. These (spatial versus non-spatial nodes) are mentioned multiple times in the rest of the methodology, so they should be clearly defined. Additionally, it is not clear whether the probability values within the nodes in the figure are showing values for a specific location/pixel, or what the significance of these values is, if any. Are they just serving as an example? Do the values indicate anything of importance? If not, perhaps they can be removed for clarity of the flow chart.

Sections 2.4-2.8 – The details of the functions used in data processing (e.g., lines 202-203, 206-210, 226, etc.) may not be necessary in the main text of the document – I would suggest moving such details to an Appendix and leaving just an overview of the data processing steps in the main text, unless this is standard practice for this journal.

Section 2.7 – Please define what is meant by “pixel state.” Otherwise, the description in Sections 2.7.1-2.7.4 is somewhat confusing.

Section 2.8 and Section 3.5.2 – A map showing the locations of validation polygons could be useful, either in the text or in an Appendix (or as an overlay on the results predicted by the model, Figure 10). This would show at a glance the geographic distribution of the samples used in model validation.

Section 3.1 – What are the main takeaways of this section, and what significance to they play in relation to the other results being reported?

Section 3.5.2, Table 4 – It seems interesting that very few pixels fall into the “Medium” FBFS potential category, with the majority categorized as either very low/low or high/very high. Is there any explanation for this?

 

Comments on writing/formatting:

Section 2.3, Lines 176-194 – This paragraph is very dense and difficult to parse without multiple readings. If possible, please edit the text for easier reading by breaking up complex sentences and simplifying the wording where possible, and add definitions of terms where they could help clarify meaning. As an example, the sentence “We assumed that the conditions of the pixel must allow the possibility for seasonal vegetation without omitting the possibility that the variation in vegetation can either be due to flood or non-flood water” could be written more simply as follows: “We assumed that vegetation within a pixel can vary seasonally either with or without flood water” (or “…in the presence of absence of flood water”).

Section 2.7 – Why are certain phrases written in bold font?

Author Response

Thank you for taking time to review our manuscript. Attached our response to your comments.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I thank the authors for the thorough and detailed response to my comments. The methods and results in the revised manuscript are much clearer and more meaningful. I especially appreciate the addition of Figure 12 - it greatly improved the interpretability of the results and method.

I am satisfied with the current revision of the manuscript and have no further edits or comments.