Thread:
@nytimes has carried an article how cities would be under water which shows #Mumbai being wiped out
We need to exercise caution before using this data n inferences in Indian context!
nytimes.com/interactive/20…
FYI
@PrinSciAdvGoI @dishashetty20 @rapiduplift @RyanMaue
@nytimes has carried an article how cities would be under water which shows #Mumbai being wiped out
We need to exercise caution before using this data n inferences in Indian context!
nytimes.com/interactive/20…
FYI
@PrinSciAdvGoI @dishashetty20 @rapiduplift @RyanMaue
Currently elevation data like SRTM have a vertical error i.e height data created by satellites don't represent ground height very accurately due to buildings, trees etc. So while doing analysis there's a possibility of understimating areas affected by floods, storm surge etc
2/n
2/n
Inorder to overcome this issue, the authors of the paper had tried to create an AI model that can correct the elevation bias in the data. The AI model is built with ~23 parameters like population density, vegetation etc to explain the variation in the satellite data.
3/n
3/n
Since building heights data is not available, population density n other parameters are used as proxies which is understandable. But the input data used are extremely coarse spatial data n don't represent ground even at metro level.
4/n
4/n
For eg: Population density data has 1 km resolution (elevation data is 30m) i.e cross-section of Mumbai would be less than 4 pixels. The errors in these creep into model results also.
5/n
5/n
Apart from this, the model was trained and tested with LIDAR Datasets from US and Aus which was later used to predict elevation bias in India and other countries which lack LIDAR data. But, the underlying parameters for the model don't behave in same way for India and US.
6/n
6/n
For eg, moderate population density in US can lead to higher buildings but not necessarily the same in India. The ratio between population density and building heights are not the same for the two countries. So it would likely overestimate the vertical bias
7/n
7/n
Due to these issues, The Kolkata Airport whose run way is located at a height of 7m from MSL and far inland has been projected to be affected by 1-2m rise in sea level!
8/n
8/n
In sensitive subjects like this, one has to make sure that ground truthing is done to make sure that the results are reliable. However in this case, it is NOT done
9/n
9/n
Since authors had difficulty in doing ground truth in India or other Asian countries, writers should have refrained making sensitive statements that cities would be wiped out etc. Authors cud have waited for ground truth to improve data rather than sending alarmist messages
10/n
10/n
Hence, the document, the methods, and the data have to be used with extreme caution when used in local context. It's better to do ground surveys for projecting the risks in these places. Alarmist false positives are as bad as false negatives.
11/n
11/n
I understand the need for usage of "Bathtub" analysis which don't take into account our ocean dynamics due to computation limitations. But we need to realize that the risks faced by communities in the coastal region are very different and are not explained by these.
12/n
12/n
In long-term to overcome the problem with the elevation bias, Governments need to do detailed mapping exercises for our coastal region which would help us prepare ourselves. ML algorithms with these parameters are not the solution..
13/n
13/n
And my opinion as published in India Today
