Classifier sand color option

[KatKonst]

Hi Kilian,
I am testing different classifier models and am curious as to how you created the classifier models for the different sand colour options. I am guessing the "default" options is the one derived directly from training the classifier as described in your papers, but I'm not quite sure how the other two models were derived.

This is a low priority question but it would be great if you found the time to get it...

Cheers
Kat

[kvos]

hi kat,
I trained the classifier with different training datasets, for example the 'bright' sand, I digitsed pixels on beaches with extremely white sand. You can create training data and re-train de classifier following the instructions here https://github.com/kvos/CoastSat/blob/master/doc/train_new_classifier.md

[KatKonst]

Hi Kilian, Thanks for getting back to me. I am currently doing precisely that: retraining the classifier and assessing the resulting models. My area of interest is the SW of England: lots of very dissipative beaches! Even though I would not say the sediment is not very light in colour, the "bright" classifier performs better. What you say makes sense, so now I have a couple more questions 😎: 1. Would I be right in assuming that the training dataset that is available in zenodo and figshare are the ones used to train the "default" versions (trained on Collaroy, Byron, Newcastle, Sawtel and Motuya), and the "bright" and "dark" models were trained on different beaches? 2. What made you decide to train different classifiers for Landsat but not S2? 3. What made you decide to train the Landsat classifier on L8 only? 4. (last one, I promise)... When I get to evaluating the new classifier, I get an "invalid input type" error (see attached pdf). Again, this is a low priority issue but if the problem is something obvious, it would be nice to sort it out. I'd like to thank you for all your help and for making CoastSat available in the first place! The number of users it already has, is a testament to how useful a tool it is! Cheers, Kat.

…

________________________________ From: Kilian Vos <notifications@github.com> Sent: Saturday, February 6, 2021 3:18:49 AM To: kvos/CoastSat <CoastSat@noreply.github.com> Cc: Aikaterini Konstantinou <a.konstantinou@plymouth.ac.uk>; Author <author@noreply.github.com> Subject: Re: [kvos/CoastSat] Classifier sand color option (#176) hi kat, I trained the classifier with different training datasets, for example the 'bright' sand, I digitsed pixels on beaches with extremely white sand. You can create training data and re-train de classifier following the instructions here https://github.com/kvos/CoastSat/blob/master/doc/train_new_classifier.md<https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.qkg1.top%2Fkvos%2FCoastSat%2Fblob%2Fmaster%2Fdoc%2Ftrain_new_classifier.md&data=04%7C01%7Ca.konstantinou%40plymouth.ac.uk%7Cb0346570112345d70d6408d8ca4decd7%7C5437e7eb83fb4d1abfd3bb247e061bf1%7C1%7C0%7C637481783323112672%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=NwLilIiZD8H2Odj2QY0Xw2NVtvrGMwqVtJmLklm1G24%3D&reserved=0> — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.qkg1.top%2Fkvos%2FCoastSat%2Fissues%2F176%23issuecomment-774388782&data=04%7C01%7Ca.konstantinou%40plymouth.ac.uk%7Cb0346570112345d70d6408d8ca4decd7%7C5437e7eb83fb4d1abfd3bb247e061bf1%7C1%7C0%7C637481783323122667%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=AS94yQFoQJE%2FeyFsc15hZWgzmxb0xWimWLsKbACqibE%3D&reserved=0>, or unsubscribe<https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.qkg1.top%2Fnotifications%2Funsubscribe-auth%2FAMFFAT6OF2GUFCTPSMFZPS3S5SYJTANCNFSM4XCZHRDQ&data=04%7C01%7Ca.konstantinou%40plymouth.ac.uk%7Cb0346570112345d70d6408d8ca4decd7%7C5437e7eb83fb4d1abfd3bb247e061bf1%7C1%7C0%7C637481783323132659%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=F%2FodQgnkTUNmB2VzoCcypDwIeOGI3eIW%2BmDZkq%2FmwF8%3D&reserved=0>.

________________________________ [http://www.plymouth.ac.uk/images/email_footer.gif]<http://www.plymouth.ac.uk/worldclass> This email and any files with it are confidential and intended solely for the use of the recipient to whom it is addressed. If you are not the intended recipient then copying, distribution or other use of the information contained is strictly prohibited and you should not rely on it. If you have received this email in error please let the sender know immediately and delete it from your system(s). Internet emails are not necessarily secure. While we take every care, University of Plymouth accepts no responsibility for viruses and it is your responsibility to scan emails and their attachments. University of Plymouth does not accept responsibility for any changes made after it was sent. Nothing in this email or its attachments constitutes an order for goods or services unless accompanied by an official order form.

[KatKonst]

New_classiffier_evaluation_error.pdf

[kvos]

hi Kat,

1. The training data used for the default classifier, both L8 and S2 is contained in the github repo in these pickle files https://github.com/kvos/CoastSat/tree/master/classification/training_data. I think it's the same as the zenodo data but the format may have changed so better to use the files from the repo as they are the most updated.
2. There's no reason for not creating different classifiers for S2 other than not having enough time ahaha
3. So Landsat sensors were designed to be inter-comparable and suitable for time-series analysis, so when they are calibrated to Top-Of-Atmosphere, the reflectance values from L5, L7 or L8 for the same object should match. That was my initial reasoning for only training on L8 (easier to digitise pixels) and using the same classifier for the other two satellite. Having said that, I have noticed some inconsistencies, especially with the 'water' class on Landsat 7 images, and I think it could be caused by the fact that pansharpened bands are given as input for the classifier and L8 and L7 have a different panchromatic band. Let me know if you need more info on this.
4. It looks like it's an issue with the shoreline detection, which is included in the evaluation function. You can go into the code of the evaluate_classifier() function and comment out all the lines that do the shoreline detection (577-598 & 615-616) as you only want to see the output of the classification (mapping the shorelines was just an extra). My guess from the error message is that the contours_mwi variable may be empty for some reason.

thanks, also it would be great if you could have a look at this #177 and maybe provide some of your observations as well

[KatKonst]

Hi Kilian,
Thank you for your reply.

1. Thanks, I'll check it out.
2. Not enough time: tell me about it!
3. I thought so. L8 B08 (panchromatic) is much narrower compared to L7 and does not include the NIR (L8: 0.5 - 0.68 μm; L7: 0.51 - 0.9μm). So I think that would explain discrepancies, especially over water. YES, I would like a bit more on that if possible.
4. Thanks.
5. I hope my comments to Wet/dry sand limit instead of the water/sand limit #177 are useful. It is a problem inherent to this type of beach environment that also baffled me for some time.

-Kat

[kvos]

3. Yes that's probably the main issue with reusing the same classifier for L7 and L8. Currently, in the CoastSat pre-processing steps, the pan band is used to pansharpen B-G-R on L8 images and G-R-NIR on L7 images, as these are the bands that are covered by the respective pan bands. Now I don't know if that is actually the state of the art in terms of pansharpening L7 and L8 collections for time-series analysis. I will ask my colleague Tino as he told me once that he saw some artefacts in the intensities of the pansharpened water pixels for L7.

[kvos]

problem 4. (evaluating the classifier) is fixed in the latest update #192