Guidance on how test all approaches proposed in the article

[hadikhodadadi]

Hello
I just read your article and while checking the results and testing your program, I have some questions. I would be grateful if you could guide me.

1. Has the generalization of the trained network on one region been checked in other regions?
2. How to change the dimensions of the coherence calculation window in ISCE? (for example 5 x 15 or 9 x 9)
3. How is the output from the deep learning network geo-referenced? (in order to compare with reference data)
   I know that dealing with how to generate coherence or georeferencing it is not related to your work, but I would be grateful if you could help in this matter.

[olliestephenson]

Thanks for your interest! Some quick responses here: 1. We have not tested the generalization from one region to another region. This is because of the way the method works – we are not training on actual damage data, but rather time series data from before the disaster. This means that, if there is available data from before the event, we will always be able to retrain the RNN on that data. It is likely a good idea to retrain, as each region will have different coherence characteristics through time. It would be interesting to train on many different locations simultaneously, and see if that provides better results (I suspect that would help avoid overfitting) 2. In ISCE, you need to alter the number of “looks” in range and azimuth direction (e.g. look at the input parameters “range_looks” and “azimuth_looks” in here: https://github.com/isce-framework/isce2/blob/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack/FilterAndCoherence.py. This is part of the topsStack processing work flow, which has a readme here: https://github.com/isce-framework/isce2/tree/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack) 3. For georeferencing, you can use the “gdal” library. ISCE has a geocoding script called “geocodeGdal.py” https://github.com/isce-framework/isce2/blob/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack/geocodeGdal.py It has been a while since I’ve used ISCE, so you may want to ask questions to the developers if you’re having issues. Good luck! From: hadikhodadadi ***@***.***> Date: Monday, May 27, 2024 at 8:39 AM To: olliestephenson/dpm-rnn-public ***@***.***> Cc: Subscribed ***@***.***> Subject: [olliestephenson/dpm-rnn-public] Guidance on how test all approaches proposed in the article (Issue #4) Hello I just read your article and while checking the results and testing your program, I have some questions. I would be grateful if you could guide me. 1. Has the generalization of the trained network on one region been checked in other regions? 2. How to change the dimensions of the coherence calculation window in ISCE? (for example 5 x 15 or 9 x 9) 3. How is the output from the deep learning network geo-referenced? (in order to compare with reference data) I know that dealing with how to generate coherence or georeferencing it is not related to your work, but I would be grateful if you could help in this matter. — Reply to this email directly, view it on GitHub<#4>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHEEFZEVQM5H7267AQ2FLY3ZEMSQZAVCNFSM6AAAAABILD62X6VHI2DSMVQWIX3LMV43ASLTON2WKOZSGMYTSMBWGQ3TGOA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[hadikhodadadi]

Thank you very much for your reply Regarding the second point, I had a question. Did you consider only the multilook changes to create coherence that lead to pixel size changes? The issue I was considering is changing the dimensions of the window used to calculate the coherence in each pixel, which is set to 5 x 5 by default in the software and is displayed under the title of correlation weights. Have you checked its changes?

…

On Wed, Jun 26, 2024 at 3:03 AM Oliver Stephenson ***@***.***> wrote: Thanks for your interest! Some quick responses here: 1. We have not tested the generalization from one region to another region. This is because of the way the method works – we are not training on actual damage data, but rather time series data from before the disaster. This means that, if there is available data from before the event, we will always be able to retrain the RNN on that data. It is likely a good idea to retrain, as each region will have different coherence characteristics through time. It would be interesting to train on many different locations simultaneously, and see if that provides better results (I suspect that would help avoid overfitting) 2. In ISCE, you need to alter the number of “looks” in range and azimuth direction (e.g. look at the input parameters “range_looks” and “azimuth_looks” in here: https://github.com/isce-framework/isce2/blob/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack/FilterAndCoherence.py. This is part of the topsStack processing work flow, which has a readme here: https://github.com/isce-framework/isce2/tree/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack) 3. For georeferencing, you can use the “gdal” library. ISCE has a geocoding script called “geocodeGdal.py” https://github.com/isce-framework/isce2/blob/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack/geocodeGdal.py It has been a while since I’ve used ISCE, so you may want to ask questions to the developers if you’re having issues. Good luck! From: hadikhodadadi ***@***.***> Date: Monday, May 27, 2024 at 8:39 AM To: olliestephenson/dpm-rnn-public ***@***.***> Cc: Subscribed ***@***.***> Subject: [olliestephenson/dpm-rnn-public] Guidance on how test all approaches proposed in the article (Issue #4) Hello I just read your article and while checking the results and testing your program, I have some questions. I would be grateful if you could guide me. 1. Has the generalization of the trained network on one region been checked in other regions? 2. How to change the dimensions of the coherence calculation window in ISCE? (for example 5 x 15 or 9 x 9) 3. How is the output from the deep learning network geo-referenced? (in order to compare with reference data) I know that dealing with how to generate coherence or georeferencing it is not related to your work, but I would be grateful if you could help in this matter. — Reply to this email directly, view it on GitHub< #4>, or unsubscribe< https://github.com/notifications/unsubscribe-auth/AHEEFZEVQM5H7267AQ2FLY3ZEMSQZAVCNFSM6AAAAABILD62X6VHI2DSMVQWIX3LMV43ASLTON2WKOZSGMYTSMBWGQ3TGOA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***> — Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A3T7TTW2TCC4MDGU3BZ7VIDZJH43HAVCNFSM6AAAAABILD62X6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCOJQGIYDCNZVG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[olliestephenson]

In our work we used a 5x15 sized chip/window. Note that the size of the pixels is different in the azimuth and range direction (SLC pixels are 2x14 meters approximately for Sentinel 1). You can read about this in the supplementary materials (SIII: https://arxiv.org/pdf/2105.11544). We didn’t systematically experiment with different chip sizes, however from some brief trials we found that smaller chips could lead to worse results. As you shrink the size of the chip, you have fewer samples, so you risk having a less reliable coherence measure. However as you make the chip larger, you span a wider area, so the coherence is less representative of a specific area (e.g. your chip might be much larger than the building which has collapsed). From: hadikhodadadi ***@***.***> Date: Monday, July 1, 2024 at 3:54 AM To: olliestephenson/dpm-rnn-public ***@***.***> Cc: Oliver Stephenson ***@***.***>, Comment ***@***.***> Subject: Re: [olliestephenson/dpm-rnn-public] Guidance on how test all approaches proposed in the article (Issue #4) Thank you very much for your reply Regarding the second point, I had a question. Did you consider only the multilook changes to create coherence that lead to pixel size changes? The issue I was considering is changing the dimensions of the window used to calculate the coherence in each pixel, which is set to 5 x 5 by default in the software and is displayed under the title of correlation weights. Have you checked its changes?

On Wed, Jun 26, 2024 at 3:03 AM Oliver Stephenson ***@***.***> wrote: Thanks for your interest! Some quick responses here: 1. We have not tested the generalization from one region to another region. This is because of the way the method works – we are not training on actual damage data, but rather time series data from before the disaster. This means that, if there is available data from before the event, we will always be able to retrain the RNN on that data. It is likely a good idea to retrain, as each region will have different coherence characteristics through time. It would be interesting to train on many different locations simultaneously, and see if that provides better results (I suspect that would help avoid overfitting) 2. In ISCE, you need to alter the number of “looks” in range and azimuth direction (e.g. look at the input parameters “range_looks” and “azimuth_looks” in here: https://github.com/isce-framework/isce2/blob/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack/FilterAndCoherence.py. This is part of the topsStack processing work flow, which has a readme here: https://github.com/isce-framework/isce2/tree/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack) 3. For georeferencing, you can use the “gdal” library. ISCE has a geocoding script called “geocodeGdal.py” https://github.com/isce-framework/isce2/blob/ca7649af691961e7330ecb8451251144893d8768/contrib/stack/topsStack/geocodeGdal.py It has been a while since I’ve used ISCE, so you may want to ask questions to the developers if you’re having issues. Good luck! From: hadikhodadadi ***@***.***> Date: Monday, May 27, 2024 at 8:39 AM To: olliestephenson/dpm-rnn-public ***@***.***> Cc: Subscribed ***@***.***> Subject: [olliestephenson/dpm-rnn-public] Guidance on how test all approaches proposed in the article (Issue #4) Hello I just read your article and while checking the results and testing your program, I have some questions. I would be grateful if you could guide me. 1. Has the generalization of the trained network on one region been checked in other regions? 2. How to change the dimensions of the coherence calculation window in ISCE? (for example 5 x 15 or 9 x 9) 3. How is the output from the deep learning network geo-referenced? (in order to compare with reference data) I know that dealing with how to generate coherence or georeferencing it is not related to your work, but I would be grateful if you could help in this matter. — Reply to this email directly, view it on GitHub< #4>, or unsubscribe< https://github.com/notifications/unsubscribe-auth/AHEEFZEVQM5H7267AQ2FLY3ZEMSQZAVCNFSM6AAAAABILD62X6VHI2DSMVQWIX3LMV43ASLTON2WKOZSGMYTSMBWGQ3TGOA>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***> — Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A3T7TTW2TCC4MDGU3BZ7VIDZJH43HAVCNFSM6AAAAABILD62X6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCOJQGIYDCNZVG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

— Reply to this email directly, view it on GitHub<#4 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHEEFZDDBWGJNZYJGQLVKJDZKEDKHAVCNFSM6AAAAABILD62X6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCOJZGQ4DANBXG4>. You are receiving this because you commented.Message ID: ***@***.***>