High-Level Project Summary
The interferometric synthetic aperture radar (InSAR) singnals have noised caused by natural phenomena including rain, snow, wind, and seasonal changes. The coupled effect of major events and natural phenomena sometimes leads to misinterpretation of interferometric coherence maps and often degrades the performance of change detection algorithms Therefore, denoising techniques are introduced in attempt to denoise these signals, get corrected information from themThe solution have been developed in three different approaches :-1- using denoising filters for images 2- using deep learning denoising model 3- using pretrained model desinged for insar specefically
Link to Final Project
Link to Project "Demo"
Detailed Project Description
The interferometric synthetic aperture radar (InSAR) singnals have noised caused by natural phenomena including rain, snow, wind, and seasonal changes. The coupled effect of major events and natural phenomena sometimes leads to misinterpretation of interferometric coherence maps and often degrades the performance of change detection algorithms Therefore, denoising techniques are introduced in attempt to denoise these signals, get corrected information from them
The solution have been developed in three different approaches :-
1- using denoising filters for images
In this approach, several filters were used to denoise images (Gaussian filter, medain filetering, bilateral filtering,
Non-local mean filtering, Total variation flitering, Block matching and 3d filtering)
2- using deep learning denoising model
In this apprach a deep learning auto encoder model to denoise image
model was trained on clean data provided by kaggle
then when a noised image is fed to the model , then the denoised image is then predicted
3- using pretrained model desinged for insar specifically
In this approach we used pretrained model to preprocess and denoise model
the pretrained model used is small_base_line_app used with mintpy library
It's hoped that this project is turned to a fully implemented UI app ready to be in the market where noised data is entered as input and the output is the denoised data
Tools used :-
colab notebooks
kaggle notebooks
pychram
languages used :-
python
libraries used :-
opencv
sklearn
tensorflow
mintpy
Space Agency Data
we used Alaska Satelite facility (ASF) data search web site to generate our dataset.
This data was in cdf format however it was our first time to deal with such format so it took us much time to know how to deal with it finally we attempt to transform it to hdf5 format where we can use and process the InSAR images.
Another data we used is insar kaggle dataset which is a prepared clean dataset initially in hdf5 format which helped us processing and denoising InSAR images directly.
Hackathon Journey
Our Space Apps experience was amazing yet worth it. We have learned many things in both the technical and social parts. In the technical part, we have learned more about ML and filters in python, how to use them and what they particularly affect. The social one was the best! All team members engaged in a robust exchange of ideas to come up with the best solution. Starting with choosing the challenge, and ending by submitting the final project.
Our approach was to differentiate decorrelation sources caused by natural changes from those caused by an event of interest, first signals received need to be processed, then formulated a temporal decorrelation model that accounts for signals separating, namely ground deformation and tropospheric noise. After the signal processing and decorrelation model, typically noise reduction techniques are used to denoise the signals.
References
https://youtu.be/_vWWAmJFQh8
https://youtu.be/1k5R4d0gb9U
https://youtu.be/M8GRcQL5N1M
https://youtu.be/R-t2utzo7mg
https://youtu.be/oKy0IAxYuaI
https://github.com/insarlab/MintPy-tutorial
https://youtu.be/W-qyvXIIaxU
https://youtu.be/m4G0zzjBpfk
Tags
#software #satelites #Insar #AI #image_processing #deep_learning #image_denoising