Automated Inundation Mapping Over Large Areas Using Landsat Data and Google Earth Engine

Round 1

Reviewer 1 Report

This manuscript describes a method to extract inundation information from 30-meter imagery from Landsat over a 30 year period. They accurately identify the spatial fluctuations of the flood variance over time with respect to previous studies in the region and identify maximum and minimum flood years. The document is well-written, and the figures are appropriate. Overall, I think this is a great paper although I believe it requires a more robust literature review, some re-writing, and figure updates. Therefore, I would recommend this paper be accepted after a major revision.

I have identified three overarching themes that I believe require more work:

• Algorithm:
  • How does this work differ from Wolski’s, with the exception of applying the formula to a new dataset?
  • Lack of presented knowledge about background literature in inundation/flood mapping at fine and coarse resolutions
    • Missing literature on vegetation and water indices (NDWI, NDVI, NDPI)
    • See section 1.1 in Wolski
• Figures
  • Do the figures accurately display the message you are trying to convey?
    • Quality/resolution and color of the images,
    • Figures about the low flood levels
    • Figures about the thresholding
      • Also, figure 2 in Wolski -- could this study plot the frequency distributions and mark the median lines for the dry/wet regions for each year (showing how those thresholds change over time)
• Organization and content priorities
  • What is more important, Okavango inundation or mapping methodologies?
    • Both have a place in the paper, but it is important to establish a clear priority of one over the other.
      • It seems as if the inundation is more important than the methodology, due to the lack of the discussion on how the method was adopted and modified for this application (there are discussions on the relevance of the spatial resolution, but there are no real comparisons between Landsat and other satellite data.)

* The claim is that you’ve have implemented a method to classify inundated regions, but with a case study of the Okavango Delta. In actuality, you were first looking into the delta hydrology and therefore the paper should prioritize the relative importance of mapping the delta over inundation mapping in general. (In fact, the current citation list shows a bias towards Okavango Delta mapping papers and there are few papers on inundation mapping techniques in general.) 

* Reorder the first few paragraphs to reflect the re-prioritization: Paragraph 2, Paragraph 3 (mix paragraph 1 in with 3).  Start first talking about the Okavango Delta, location and wetland importance, then move to complex hydrology and mapping methods and the importance of inundation maps.

~

Abstract- Line 18: On average how much smaller is the inundation extent? Compared to what other methods, MODIS? Is the percent difference in inundation extent topography driven or seasonal?

Line 29: Original sentence: “If the area of interest is large or difficult to access, if multiple maps are required, or human resources are limited, then inundation maps can be created using a historical suite of satellite imagery.” This is a little bit confusing try something like** “Inundation maps can be created using satellite images, which are available from a range of spatial and temporal resolution products. High spatial resolution information can increase confidence in associated decision-making.”

Line 55: ‘but see [4, 16]’ Is this a typo?

Line 56: ‘Sub-monthly analysis’… There are also studies at the daily temporal resolution that downscale coarse resolution MODIS data.

Daily MODIS Inundation Literature:

* Islam, A. S., Bala, S. K., & Haque, M. A. (2010). Flood inundation map of Bangladesh using MODIS time‐series images. Journal of Flood Risk Management, 3(3), 210-222.

* Chen, Y., Huang, C., Ticehurst, C., Merrin, L., & Thew, P. (2013). An evaluation of MODIS daily and 8-day composite products for floodplain and wetland inundation mapping. Wetlands, 33(5), 823-835.

* Ticehurst, C., Dutta, D., Karim, F., Petheram, C., & Guerschman, J. P. (2015). Improving the accuracy of daily MODIS OWL flood inundation mapping using hydrodynamic modelling. Natural Hazards, 78(2), 803-820.

*Fayne, J. V., Bolten, J. D., Doyle, C. S., Fuhrmann, S., Rice, M. T., Houser, P. R., & Lakshmi, V. (2017). Flood mapping in the lower Mekong River Basin using daily MODIS observations. International journal of remote sensing, 38(6), 1737-1757.

Line 59: ‘mixed pixels’ See ‘OWL’ papers (and others) to learn how coarse resolution data is used in inundation mapping

Line 59: ‘mixed pixels’ here you define mixed pixels (in this particular context), although this is not universally true for mixed pixels. Can you explain why you’ve selected this threshold and what that assumption means for the remainder of your study? Also, can you briefly define what you mean by spectral overlap in-text, and how does the proposed methodology solve this problem?

Line 82: ‘classed’ --> ‘classified’

Line 82: ‘cfmask’ --> ‘Landsat cloud mask’

Line 84: ‘pre-existing algorithm’ à ‘cloud-masking and gap-filling algorithm’

Line 84: stackoverflow link à replace with a proper citation

Line 86: What does it mean for a whole scene to be missing? Does that mean that a single date observation was completely cloud covered and therefore that year has to be omitted from the analysis? How frequently does this happen?

The dates (1993, 2000, 2009, 2010, and 2012) are revealed at the end.

Also, is this really necessary? How much variance does each input scene contribute to the variance that the opinion is 'all or nothing' with regard to retaining a year?

Figure 1: It is strange to mark a river channel as 'permanently inundated'. If water is present consistently, can it be considered inundation? Or are these areas only ‘permanently inundated’ during this flood season?

Eq. 1: How does this differ from Wolski’s equation? In Wolski, f is different values but the recommended default is 0.3 (for 50%) as assessed for coarse resolution imagery.  I assume you are using the f=0.3 because the pixels are 50% (section 3.3 in Wolski). Please explain why you chose a 50% mixing ratio for this study.

Figure 2a: While this might technically be 'true color' for this region, many readers expect to see mixtures of color (blue/green/black for water, green/brown for land) try using a false color composite which might help features become more easily distinguishable. For the greyscale SWIR composite, use a histogram stretch or hill-shading to make features more visible. 

Figure 2b: ‘showing the fine detail possible’ à awkward grammar, just delete; Landsat imagery detail does not look like anything when it’s not compared with a lower resolution image. (So maybe add in a MODIS image for reference).

Line 113: ‘true color versions of Landsat composites’. Again, it is difficult to rely on our eyes when 'true color' does not produce easily recognizable features. This is an even more valid point in a braided channel with mixed pixels. It is important that we feel confident in the ability to visibly assess the imagery if the visual assessment will be a primary form of validation.  Make sure to use color composites that make the flood area stand out.

Line 119-124: This section is a little bit confusing so my assessment might be incorrect here. If so that means this section needs to be re-written for clarity.

Line 119-124: 50 sample points were assigned 'wet or dry' before progressing to the next year. Do you use the same 50 points for all years or do you generate a new set of 50 points? (Is this what you meant by avoiding bias of temporally proximate samples?)  To reduce confusion here, you could produce a map of the points (S1B) showing the frequency each point was correctly classified over the 30 years. Here, you can identify if your error is distributed evenly throughout the delta or if there is a localized problem.

Line 126: Suggested re-write: To validate the accuracy of the classification method, we carried out a field examination of the inundated regions within one Landsat scene (xx)

Line 128: Suggested re-write: Due to accessibility and safety constraints, we only sampled from the Abu Concession [F-S2] where inundation was shallow enough for field personnel to access by wading, within 100 meters of dry land.

Line 146: Missing inundation map years: You originally mention this in 'materials and methods'. If the compositing was a 'method' then it makes sense for the compositing to be a result. In my view, the compositing was the background to the method (the real method being the thresholding SWIR values) and therefore the composite is just a material. In that case, this sentence (missing map years) belongs in the materials section so we know right away how many years are missing.

Line 149: The Landsat 7 sensor malfunction occurred in 2003. You also used Landsat 5 (and 8, but 8 had not launched by 2011). Was Landsat 5 not enough to cover the gaps in Landsat 7? 

Figure 4: Include the sensors used or product spatial resolution.

Line 162: Typo: ‘validations pixels’à ‘validation pixels’

Line 178: Suggested re-write: 'Based on the maps produced in this study, the flood event represents the smallest inundation since 1985...'

Line 178: Here, we come back to your key idea: the ability of inundation maps inform water resources research and management in the region.

Figure 3 and 4 are close to reaching this point, but you want to also describe the range of spatial variability in the delta.

A proposed Figure 5 might take the composite image and derived inundation maps for three time periods (min, mean, max inundation -- 2019, ?, 2011)

(2x3 figure of true (or false) color images and derived inundation maps showing the spatial variability of the flooding)

This also gives us another visual representation of how well the inundated areas are classified.

Line 180-182: Simplify this sentence, the years make it confusing.

*Also those researchers did not 'predict' a past extent. This is an estimate.

Line 180-182:  Suggested re-write: 'Estimations of inundation extent going back to 1934 calculated the lowest inundation to be xx km'. Our estimates of inundation extent in 2019 was xx km, making it the smallest flood in 85 years.'

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Summary:

This technical note presents an adaptation of coarse-resolution (MODIS) remote sensing inundation mapping technique to a fine-resolution sensor (Landsat).  It is novel because it extends the existing time series of inundation maps for the Okavango Delta (OD).  The work was done using 30m Landsat imagery in GEE, and its “high” spatial resolution is also novel.  The chosen thresholding method was used to create annual peak water season inundation maps for the OD, validated with visual inspection of two scales of imagery and from field assessment.  The subsequent inundation area time series has the same general shape as the model study and others.

Overall comments:

This manuscript presents well-thought-out and methodologically sound research.  I noted one area where I thought the error analysis could have been improved. Even so, your overall accuracy seems highly believable.  If you choose to dig deeper in your discussion section, I would have liked to hear more about the anomalously-low flood year of 2019, perhaps instead of concluding remarks on the accuracy assessment.  It seems like yours will be the first remote sensing study to capture this anomalous event in the OD, and it will probably be cited for this.

The supplementary figures would benefit from being combined into a pdf document with captions.  For example, it is unclear if the grey, inundated area shown in both figures is maximum inundation, or an annual snapshot.  The supplemental rasters look very believable- congratulations on this useful product!

Specific comments:

44 paragraph:    Although it is simple, I do believe that band thresholding can provide accurate results.  Still, can you add another sentence or two justifying the use of this technique, beyond just stating the accuracy assessment results (i.e. less expensive, not necessary if the land is homogenous)?  Many water mapping papers use more sophisticated techniques.  Perhaps you can point to the band indices investigated in Wolski et al. 2017.

93         About how much would the total inundated area change if you slightly changed the SWIR threshold?  Could this explain your lower overall areas from the Wolski et all 2017 MODIS study?

93-95    Your permanently dry polygons seem to only comprise the desert landcover type.  Would including denser, dry vegetation areas better differentiate your final classes, or does your method work better with two, homogeneous end members?

108       Section 2.3 describes your image-based accuracy assessment, but you don’t restrict your sampling to the water magins, as you do for the in-situ assessment.  From Figure S2, it looks like not many of these random points fall in regions that might have annual inundation differences, thus providing for an “easy” standard with which to compare.  I think this visual analysis would be stronger if you had more aggressively clipped out the Kalahari desert/ dry regions or even parts of the permanently-inundated channel.  Fortunately, due to your solid in-situ assessment, I wouldn’t expect you to re-do the accuracy analysis.

125       This method seems very sound, and even though the water-margin points were chosen for safety and access reasons, they also would be  the hardest to classify, and thus represent a good, conservative benchmark from which to compare the classification.

209-211             “it represents a rarely conducted true

accuracy assessment of delta inundation mapping. In addition, given the sampling area was centered

on small islands and edges of the floodplain, it fittingly represents the boundary between dry and

inundated areas, the area where most classification errors are likely to occur” <- agreed!

221       As I also noted earlier, do you have a sense from your work or the literature (e.g. Wolsi Fig. 5) how much of a difference in area this could amount to?  Would it affect the relative differences between years, or just present as a shift in overall area magnitude for the whole time series?

Here are some non-essential grammar and writing suggestions:

37         Change to “flood-pulsed,” perhaps would make this sentence more concise.

46         “on satellite imagery” is unclear

51         Change “trialled” to something else, like “implemented” – or maybe this is an English dialect technicality

71         “Flood season” instead of “flood” would be clearer and doesn’t sound like a paradox

173       Should read “values in parentheses”

231-233             This sentence sounds like repeat from the intro, and can probably be excluded.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper proposed an automatic method for classifying inundation using Landsat imagery and Google Earth Engine. Then they demonstrated the method in the Okavango Delta with a classification accuracy of over 90%. As shown in abstract, the innovation of this paper can be summarized as follows: (1) presented an automatic method using GEE to extract inundation maps; (2) the spatial resolution (30m) is higher than previous studies; (3) provided longest time series (1990-2019). However, these points are nothing new and it can not meet the requirements of the high-ranking of Remote Sensing. There are many major concerns in the paper that need to be revised.

1. the SWIR thresholding technique is really too simple. I think it is very difficult to extract inundation areas automatically by this method. In addition, the result extracted by the present method is water body maps not inundation maps. In order to extract the fine inundation areas, you’d better use a change detection method. First of all, you should calculate the region of water body at the normal level, and then the region of the flooded region can be extract.
2. Another reason why I do not trust the method of this article is that the areas of water body are greatly affected by image acquire time. For example, there is a big difference in the water level between the rainy season and the dry season. Therefore, it is necessary to strictly select the time of the image before producing inundation maps.
3. In validation section, the results from 2000 to 2016 are evaluated, so how to evaluate the results from 1990 to 2000? In addition, the current validation methods are too subjective, which leads to high accuracy evaluation results.
4. Since a classification method has been proposed, the authors should select more experimental areas to test the applicability of the method.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Major remarks of the review may be summarized as follows:

1)  It seems that in section 2.1 (Annual (July-September) Landsat composition) for the sake of clarity a more detailed description of the applied method should be presented. For instance, the text can be additionally illustrated with a schematic diagram showing the course of action.

2) In section 2.3. and 2.4, the errors matrix (describing accuracy of inundation maps) should be defined by means of appropriate formulas.

Minor remarks:

1) In lines 118 and 119 the references to figure (Figure S1) are incorrect.

2) Similarly, in lines 129 and 135 the references to figure (Figure S2) are incorrect.

3) It seems that Figs. 3 and 4 are too large.

Concluding, the manuscript requires a minor revision before publication in Remote Sensing MDPI.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors produced a fast turn-around of updated and polished figures, more detailed text, and added 16 citations. Great job!

Because this paper is a technical note and not a full article, the responses from some reviewers (including myself) might seem a little heavy-handed in asking for more vigorous analysis, methodological detail, and novelty. The description of the technical note says: ‘These are typically brief (less than 10 pages) explanations of a single concept, technique, or study. They contain less information than an article and are suitable for rapid dissemination of results.’ To that effect, the submitted manuscript fits the word length but is more similar to a short article in terms of the style and content. In this case, it is too short and not detailed/rigorous enough to be a full article, and does not focus on the technicality [being a technical note] of any particular aspect of the study. Because of this, it is difficult for me to submit helpful suggestions beyond my previous points. I am impressed by how quickly the researchers were able to robustly respond to the other reviewer’s comments, produce new figures and edit the text. I am excited to see the completion of this manuscript, pending minor revisions.

The manuscript presents a re-application of a technical methodology (mapping flood inundation using dynamic SWIR thresholds –Wolski et al), to a longer, higher resolution time series (Landsat series) using a new programming software and interface (Google Earth Engine).

In response to the points of novelty:

"We consider our work is novel in several ways:

1. We demonstrate its applicability to a new, large data set.
2. We automate the process and supply open source code for an open source, accessible platform, so it can easily be adapted by other researchers and stakeholders (for both the delta and other regions). Remote sensing methods are otherwise difficult or impossible for non-experts to implement.
3. We doubled the number of years the method was applied to, including some significant years for flooding in the delta (e.g. 2019).
4. We conducted two forms of accuracy assessment, including an in-situ accuracy assessment which has only once been done before on an Okavango Delta inundation study (and not in Wolski et al. (2017)).
5. We supplied the end product for public dissemination – the first time a long time-series for the delta is available to the public.

First, I would consider the points a and c to be similar since changing from using MODIS (Wolski) to using Landsat affords an increase in the number of years available, also making it a ‘new large data set’.

• It's novel because you applied the method to a new dataset

b.You automate the process and make the code available for public use. Supplying the source code is a very good thing to do, but does not add to the novelty of the science. In addition, just as ‘remote sensing methods are otherwise difficult or impossible for non-experts to implement’ so is using Google Earth Engine, so it is doubtful that non-remote sensing scientists will be able to use this without training.

d.You need to conduct an accuracy assessment (or two). This is not novel science, but a requirement of scientific publishing. The methods used for accuracy assessment were also not novel.

e.I’m a little bit confused about which end-product the researchers are referring to here. ‘The first time a long time-series for the delta is available to the public.’ Figure 4 shows quite clearly that several studies (including Wolski et al) produced a long time-series of the delta. These are the products as I see them:

• Summed Annual Inundation Map (Figure 3, Data S2?)
• Time Series of Inundation Extent (Figure 4, Data S4)
• The Google Earth Engine Code

I appreciate this study/technical note and I do not believe that it requires 5 points of novelty. One is good enough.

A small note—that you can choose to ignore—is that the supplemental figures 2, and 3 show ‘incorrectly classified’ and ‘correctly classified’. It would be useful for us to know how frequently these were incorrectly classified. In figure 2 specifically, you could use a color ramp to plot the spatial variation of those incorrectly classified. How do I know whether the points that are ‘sometimes incorrectly classified’ are incorrectly classified 15% or 85% of the time? I suspect that because there are so many correctly classified that those that are incorrectly classified would only be incorrectly classified a small percentage of the time, which would be a good finding to report along with your tables.

Going back to the discussion on Google Earth Engine: In this application, Google Earth Engine (GEE) is the tool that is used to process data. It is used like any other programming language is used and discussed the same amount. That is, the same study could (on a computing system/server that can handle the data/processing load) be conducted using other programming languages. To that end, GEE is not used or discussed in any particularly unique manner.  There is no discussion about how to use GEE that might be relevant to stakeholders who do not have expertise with remote sensing or GEE.

If that is the goal of this paper, more discussion and figures/screenshots would be needed for how to use GEE to access the flood inundation analysis. However, that seems like something outside of the scope of the paper that has been presented (and therefore a possibility for an equally-short follow-along more technical paper [In this sense of, this is what we did, this is how you do it]). Since it is probably not the goal of this paper, I would recommend the simple fix to remove the mention of GEE from the title, as the programming language that you’ve chosen to do the analysis is not immediately relevant to your analysis or discussion.  

Finally ‘inundation history mapping’ is not a commonly used term-phrase used to describe what is regularly called ‘inundation mapping’ the historical part is a given. If the historical aspect is important to you for the title, you can use the phrase “Landsat Series” and we’ll know that this work is done over multiple dates.  Suggested title: ‘Automated Inundation Mapping in the Okavango Delta using the Landsat Series’ or ‘Automated Inundation Mapping over Large Areas using Dynamic SWIR Thresholds of Landsat Data’. (These are just suggestions so you get the idea).

Great work! Best wishes.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

First of all, I would like to thank the authors for their responses to my comments, but unfortunately, I still cannot agree with this version:

1. I would like to say that the 3 novel aspects I mentioned in my previous comments are not what I think they are, but quote the views of the authors. On the contrary, I do not agree that these 3 points are innovative, these 3 points really do not have much new ideas. There has been a lot of research on using remote sensing to monitor long-time series and using higher-resolution images to monitor the inundation area (Maybe other studies are in other areas).
2. The authors did not directly answer or modify the question I mentioned in the previous version, but only made an explanation, which I think is not enough.
3. In addition, I still insist that although this paper has some value, the method itself is really too simple. Because this is a scientific paper, it needs to give some inspiration to other researchers in terms of methods (or interesting finding), rather than providing a tool for others to use (or open source code).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf