When deep learning meets satellite imagery
Vložit
- čas přidán 7. 09. 2024
- When deep learning meets satellite imagery
--
A handy guide to understanding the specificities and challenges of satellite images when using deep learning.
--
CREDITS
Speaker - Julie Imbert
Script - Julie Imbert and Ségolène Husson
With the participation of
Renaud, Tugdual, Thomas and all the Earthcube team
Filmmaker, editing & motion design - Julien Mascheroni
ILLUSTRATIONS
Intellectual property
Earthcube, proprietary detections
Satellite images
fMoW dataset, Functional Map of the World, CVPR, Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee, 2018
Maxar
Metadata
Maxar
Aerial image
USGS
ARTICLES
Sara Sabour and Nicholas Frosst and Geoffrey E Hinton, Dynamic Routing Between Capsules,2017 arXiv
Y. Lecun, L. Bottou, Y. Bengio and P. Haffner, Gradient-based learning applied to document recognition, in Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, Nov. 1998.
--
• Video
www.earthcube.eu
/ earthcube
/ improving-the-descript...
This reminds me of the way Johny Srouji presents a new Apple Silicon architecture in the Apple keynotes. Very precise, focused and polished.
Thanks Julie and Earthcube Team!
This is the best video on the subject. Please set up a Udemy course on deep learning and GIS satellite imagery.
Good ideas in the presentation. For handling the contrast of the colour, you could convert all photos to black and white for training and testing. I never considered cutting images into smaller sizes and processing them as multiple tests of the same image, I had only considered cropping images to where the objects were most likely to be. Thanks for sharing.
Deep learning seems like a perfect match for hyperspectral imaging.
Next generation frenglish + 99% of the energy spent in this video got consumed by the very expressive random hand motion :D Thanks for the content.
such a disgusting comment
Great Channel. Great Idea. Her accent sounds great for American Ears. Very attractive rhythm to listen too.
Awesome explanation! 😊 Looking forward to learning more
Precise explanation. Thank you very much
So concise and clear.
Nice Explanation, worth watching, very informative
Awesome content. Thanks!
Le vocal est très bien réussi ainsi que la présentation. J'ai beaucoup aimé
Loved this
very nice video and the articulation is very good.
Thank you for the video.
Somehow I’m a bit disappointed, but it’s my fault. From the title, I expected to hear how deep learning is complementing convolutional networks, but deep learning is mentioned just superficially. Maybe a next video?
Thanks for sharing this wonderful video
Love this video!
great presentation.. i was looking this
Multiple images for same trajectory to superimpose will give accuracy
Wonderful 👏
So many relevant information in this video. Thank you so much
i speak fluent English, but i bet it doesn't sound this smooth and beautiful!!!
Very good presentation
Wonderful video
Very useful video & it seems nice company
Thanks for the video!
Which software can provide satelite digitization with AI?
Great!
The source from witch to attract satellite imagery is limited or limited to just, say one satellite?
To have a off-Nadir is sometimes better than Nadir, for instance knowing height on a building.
Also is there any consideration of privacy...
Awesome explanation.Thank you for sharing
Hi Julie,
I listened to your CZcams video with great interest and found it extremely professional and informative...really outstanding job! I was hoping you may be able to answer a couple of questions I have that are inspired by your presentation:
Given:
- NVIDIA's plans to build its "Eos" supercomputer, a machine expected to be the world’s fastest AI system after it begins operations later this year, and its advertised specifications:
-- 18.4 exaflops of AI computing performance
-- 275 petaflops of traditional scientific computing performance
- the current state of the art deep learning algorithms as they apply to satellite images; be they Earthcube's, Preligens' or a widely accepted performance standard
- the following Worldview-3's Specs:
-- 0.31m Pan Sensor resolution (aka GSD, Ground Sample Distance; off-nadir is geometric mean)
-- 680,000 sq km capacity per day
1. How many small tiles do you estimate would have to be created in order to be processed by the current neural networks?
2. How long do you estimate it take NVIDIA's Eos to process the number of tiles estimated in question one?
3. On average, how many distinct objects are classified by the current neural networks used in commercial or laboratory settings?
4. Looking ten years out, how many distinct objects do you think neural networks, coupled with forecasted improvements in conventional (non-quantum) supercomputering, will be able to simultaneously classify?
I sincerely appreciate your time and please don't spend too much of it figuring out highly informed answers. I'm just looking for "ballpark" figures from a recognized expert in the field.
If you have a Patreon or a PayPal account, I am more than willing to entertain compensation for the time you spend answering my questions.
I wish you and yours the best, and I hope all are safe, happy, and healthy. :-)
Thank you and kind regards,
Rob
Reference: nvidianews.nvidia.com/news/nvidia-announces-dgx-h100-systems-worlds-most-advanced-enterprise-ai-infrastructure
Less hand movements would be nice ... makes me think that she's either Italian or an AI with a human fault intentional inserted and overdone to make an NPC appear more human.
... but this is pure semantics. The content she presents is top-notch.
Great Explanation, Very simply explained Thanks a lot!
Another very well designed and explained video, great job!
Great explanation and information. Thank you
This was fascinating
This is a great video. Thanks!
Good work!! thank you
Great lectures
magnifique vidéo
Where can I learn more about these topics ? Im a fourth year astronomy student and find this super interesting
do you recommned me to take planetary science or radio astronomy for a thesis project>?
@@MMAisFedor whatever you love more. In the end it all comes down to how much passion you have for the field. choose something that you love
هل يمكن الكشف عن الذهب عن طريق الأقمار الصناعية
@@ouserirouserir8011 هل هناك طريقة لكشف الكنوز اخي
thanks
what are the other channels present in the satellite images?
This was a great watch! Kudos to the whole team!
i came here after watching add and now i feel awsome
Does this have to do anything with geomatics?
النَّظِيرُ فِي عِلْمِ الْفَلَكِ : نُقْطَةٌ فِي السَّمَاءِ عَلَى خَطٍّ عَمُودِيٍّ مِنَ الْمُرَاقِبِ تَحْتَ قَدَمَيْهِ مُبَاشَرَةً
Besides it’s a good presentation, I was impressed that the westerns used the Arabic word Nadir which is exactly equivalent to the scientific word so much so they took it as is (they do not have the equivalent). How marvelous our Arabic language is.
That was very interesting.
Holy mf raw audio
i like your accent
I need a project for this please help
How she says photographee is interesting. I sometimes wonder how the hell people learn some of the random ass rules in English
I keep looking at the hand movement and forgot to listen. It is weird.
When deep learning, meets satellite imagery, what do you get ? Fires like Hawaii that can burn steel and take 90 degree corners like a street racer. Thank you I needed to see that to truly have the pieces snap into focus, no stupid puns or jokes at all. Serious as those fires.
From baking cupcakes, to this.... hmm.
هل يمكن الكشف عن الذهب عن طريق الأقمار الصناعية
@
ٴ ممكن مساعدة اخي
Where can we download the dataset of satellite image ?
Did you get an answer, I would also be interested in the dataset?
You can get it freely in usgs website
Really good content, but please change the host/speaker
Tone is Russian language in British
@Basma Dokkar I think she is Russian. I have fall in love with him.
@@abu.selim26my guess was French. But it doesn't matter much. It's clear she knows her stuff and can present it well
@@Jo-re2ye If she contact me, I would be grateful
Are u French?
Nice accent
No clue what she said.
Agreed. Surely they had someone with way less of an accent.
Fun Fact: She's AI
? she's my sister
😂
Can you see me?
произношение немного режет уши :)
а так, интересный материал
The presenter is fake. She is a robot because no one with human feelings and a soul can move their arms and hands like that. Someone who controls the robot thought it was a good idea to give the AI 10,000 hours of a conductor's time, so that it could learn to move its arms. Unreal. It's like watching a 90s robot dance, perfect movements, zero soul. She is so histrionic that it does not help in the explanation, it simply, ..., distracts. It's like a child's performance at school.