Malawi map output for observational trigger

[caldwellst]

Hello all, @joseepoirier had the very good idea to make sure we discuss this issue and finalise the decision while we still have @hannahker around! Also for @Tinkaa.

The question:

For the Malawi trigger, we have agreed to produce a map at the ADMIN3 level once the triggers have been met. The exact definition of what we will show has yet to be defined, however this footnote is included in the published Anticipatory Action Framework:

CHD will produce a map of rainfall patterns for all TAs (ADMIN3) in the Southern region. It could report rainfall received during the dates of each of the 3 dry spells, and/or since 1 January (cumulative rainfall to date). The map would provide a comparison of relative precipitation amounts (which areas received more vs less) to inform geographical targeting for implementation.

ADMIN3 issue

We are unable to use the centroid method of mapping for TAs because many TAs fail to have any centroids within them. We would have to use the touching method of aggregating raster data to TAs, which doesn't match the methods used for capturing dry spells. This is an issue for all proposed methods.

I propose that rather than aggregating to ADMIN3, that whatever the final map is, we show ADMIN3 boundaries and overlay that with unaggregated raster cells. I think this makes the map much clearer, and avoids the issue that we aggregate using a different method than we used for calculating the dry spells. Whatever your preferred option below is, keep this ADMIN3 issue in mind and how you would like to aggregate (or not) the map.

Possible options:

Cumulative since January 1

As described above, show cumulative rainfall from January 1 to the date where the trigger is activated. There are a few issues I can think of here.

• Cumulative rainfall does not necessarily have relations to observed dry spells. We are tracking dry spells because the prolonged absence of rain is a significant issue, whereas this just measures overall rainfall across a time period, which could be low levels of sustained rainfall for instance.

Rainfall during the dates of 3 dry spells

Show the rainfall observed during the 3 dry spells. The primary issue I see here is:

• Unless the 3 dry spells are across the exact same dates, it's unclear what dates we show precipitation for. If they overlap, we could show across the full range of dry spell dates, but then again very easy to have misleading maps where precipitation in an area with a dry spell could look higher than those without.
• What happens if we observe a dry spell later on in the period, such as a few weeks later? Then do we only show dry spell observed precipitation for areas that saw dry spells? For areas outside, what do we show?

Mapping where dry spells occurred

This option would just be if we were overlaying raster cells, because otherwise it's just a map of ADMIN2 dry spells but with ADMIN3 boundaries. This would just highlight raster cells that observed a dry spell across our entire period of observation. This works because the sum of averages and averages of sums are equivalent. An issue here raised by @Tinkaa is:

• It doesn't capture areas where there might've been 13 days without rainfall, for instance, or areas that saw extremely long dry spells (maybe 21+ days), so considers only the binary 14 days+ or nothing.

Mapping number of consecutive dry days

In response to @Tinkaa's point above, I would propose another option that would be mapping the number of consecutive dry days for each raster cell, measured across the entire period of observation. An issue I have is:

• In terms of data visualization, I think this becomes less clear than the binary map in terms of immediately visualizing the areas we want to respond in, as the scale difference from 12 to 14 and beyond is not going to be extremely differentiating (depends on the distribution of the data of course).

My vote

Considering the above, I would not vote at all to display precipitation, either at the raster or TA level, because I don't think that method produces a clear, understandable visual. I also think it could mislead the response because cumulative precipitation is not directly correlated with dry spells, even if there is some relation. Cumulative precipitation just during dry spells I think becomes extremely difficult to interpret (and to update) as soon as dry spells don't overlap or even differ slightly in dates. I also think we should only be displaying at the raster level with ADMIN3 overlaid.

I would vote that we do a binary mapping of raster cells and where they occur, because it makes a clear map that is easy to interpret and respond to. However, I'm a big ranked choice voting fan, so will put my votes as below.

1. Binary raster
2. # of days raster
3. Cumulative precipitation (entire period) raster
4. Cumulative precipitation (during dry spells) raster

Look forward to hearing what y'all think. There might be additional options that come up, and I will of course update my voting as those come along! Thanks!

[joseepoirier]

Thanks @caldwellst for the detailed summary. Quick points:

• The "since 1 January" parameter was inserted before we had finalised the observational trigger and decided to consider rainfall in late Dec. We could adjust the date to be since beginning of rainy season or since 24 Dec (monitoring period onset), for instance.
• Remember that the trigger's purpose is only to say WHEN not WHERE to act. The trigger, if met, will give us a list of adm2's that experienced >=1 dry spells. (Ie no need to map a binary" experienced/didn't experience a spell")
• The request for adm3 detail is to inform operations and specifically, targeting of the implementation of activities. At this point, whether or not that specific adm3 has experienced a dry spell is no longer the key question. Rather, agencies have asked to understand the current state of adm3 within an "activated" adm2 so as to inform where they will act. The implication is that there may be hum need caused by the dry spell that doesn't geographically overlap with the dry spell itself.
• Reporting total precipitation seems to me to provide that context. Precipitation during the dry spells seems somewhat of a non-starter as no precipitation will have fallen by definition.
• I do think # of days without rain could bring additional information that total precipitation doesn't capture.
• Fine by me to include the raster with an adm3 overlay rather than aggregations.

My ranked choice then would be:

1. Cum precipitation (entire period) raster + # of dry days in entire period raster (2 separate maps)
2. Sorry, not seeing enough value in the other options listed but open to discussing! :)

[Tinkaa]

Thanks for this conversation!

I would agree to do it at raster level, not aggregated.

Also, I would only show the Southern region as that is what we are monitoring.

Then, regarding the value to show.

I would agree with Seth, in that I wouldn't show cumulative precipitation. As this value relates to seasonal drought and not dry spells. I would therefore be scared that this conveys the wrong message and opens up the whole drought vs dry spell discussion again. @joseepoirier I am therefore also not fully understanding how you would say the cum precip provides the "understandment of the current status of the adm3". But I do want to understand what you mean, so could you try to explain it a bit more?

When talking about dry days, there is a distinction between # of dry days and # of consecutive dry days. I would argue the second one makes more sense since as we know a longer consecutive period of dryness has a higher impact. @caldwellst regarding the coloring, could we use a distinctive scheme where the boundary is at 14 days? I made a quick example below. So in that case all the red could mean 14+ days where darker red is more, and all the blue 14- days where darker blue is more.

I also wouldn't disregard the binary raster. As it could still better show the widespreadness of the dry spell, without the arbitrary admin2 boundaries that we are using now to determine the trigger.

So to vote I would say:

• max # consecutive dry days till trigger activation
• binary yes/no dry spell observed till trigger activation
• cumulative precipitation till trigger activation

I would agree that total precipitation during the dry spell becomes complicated

The example image (yellow is adm3 bounds, black adm2 bounds):

[Tinkaa]

Also linking #210 here as that was the start of the discussion

[joseepoirier]

Thanks Tinka for your detailed input.

I may be misunderstanding a concern Seth, you and Hannah have expressed earlier in representing or reporting dry spells at the adm3 level. But if the goal is to indicate where within a triggered adm2 the dry spell occurred and all three of you (including Hannah) are comfortable with representing the data at adm3, we can add a binary map indicating the presence of a dry spell per adm3. If you are suggesting not to map it by adm3 but instead leaving the overall raster with a binary code (ds/no ds) per cell, we can present that. Let's make sure we have a clear and convincing explanation for why we can't list adm3's with a binary DS/no DS but can represent it with sub-adm2 granularity.

The map's purpose isn't to convey where a dry spell has occurred; it is to inform on the context around those dry spells so implementation activities reflect that. That is, on what the "dryness" situation may be at the moment of activation. While I agree that it blurs the line between drought and dry spells, this is in a way no longer about whether a shock happen but about where "damage" may be greater. Cum precipitation has been requested as a proxy for what areas might be most suffering from dry conditions. Since activities should only implemented in adm2's where dry spells occurred, i think it is safe to assume that adm3's will be targeted for their "dryness" due to at least dry spells (and maybe, overall drought). Hope this helps.

Consecutive vs non-consecutive dry days: I don't have a strong preference. It seems # (non-consecutive) may be more informative because complimentary to the the statement that an area has experienced a dry spell (ie we already know they've had at least 14 cons days). But maybe that's too closely correlated with cum rainfall. Will let Seth and you look at the data and offer your recommendation.

[caldwellst]

Thanks Josée for your additional input, clear. Hannah said yesterday that she would also review and add in her comments, so am just waiting for her thoughts before responding.

[hannahker]

A super interesting conversation here and I'll mostly just echo what others have already said.

• I definitely agree that the only way to show anything meaningful at the admin 3 level is to overlay the boundaries on the raster. The resolution of the raster is way too coarse to do any zonal aggregation at the admin 3 level. To justify this, we might say something along the lines of "We can provide a map with admin 3 boundaries overlaid on the raw data to suggest which admin 3s are affected, but are uncomfortable providing a list of admin 3s as this would be providing a false indication of the precision of our source dataset."
• I'm also echoing @Tinkaa and @caldwellst's view that showing cumulative precipitation could be misleading as it's not necessarily contributing anything to our understanding of the sustained dryness that this trigger is designed to capture. I see your main argument for this @joseepoirier, but I'm worried that we run the risk of this output contributing to less trust in our identification of dry spells, as perhaps the output map might show that our triggered area has overall more precipitation than other areas. This might be a confusing message.
• My votes also go the same as @Tinkaa

For the example figure, the diverging colour scheme is a nice idea, but I'd just probably not use blue for anything as it might be misinterpreted as precipitation.