High-Level Project Summary
our project pipeline was the following: firstly we built a filtration function of humidity percentage so we can determine if this percentage need to filtered so we move to the next steps in our pipeline otherwise we will neglect the deformation of this little percentage.secondly we built a Deep Learning model (classification task- CNN) to classify the InSAR images to four classes which are: Earthquake, Volcano, Land Slide, Tectonic Movements.Third we applied some signal processing Fourth we built a convolutional auto-encoder for InSAR images denoising (so we can remove the water vapor noise from the InSAR images).
Link to Final Project
Link to Project "Demo"
Detailed Project Description
Project Pipeline:
- Built a filtration function, if the percentage of the humidity (water-vapor) is very little we will neglect its affect on the InSAR images, else we need to measure or remove its affect in the InSAR images generated.
- Then we build a classification task (VGG-16 CNN-based model) to classify the input images (InSAR images) to four classes which are: Earthquake, Volcano, Land Slide, Tectonic Movements. This step will clear to us how we can work with it in the next step and how can we apply filters, data analysis or signal processing on it.
- The third step is signal processing, We detected Amplitudes (peaks) of INSAR data from netCDF files which allocated in NASA resources for San Andreas fault, and assumed peeks of pure deformation data Which is supposed to got it from GPS stations datasets. we subtracted peaks of INSAR data from GPS data to get tropospheric noise peaks. Then we can subtracted INSAR peaks from tropospheric noise peaks to get INSAR deformation without noise.
- At the end we built an Auto-Encoder (CNN-based) to denoising the input InSAR images. I pass to this auto-encoder input images as InSAR image and the output will be the same input image without the humidity (water-vapor) deformation (noise), So auto-encoder will trained and learned or generated the transformation matrix which in production (testing) will applied in the new InSAR data and we get it with out the deformation or noise or error due to the water vapor in atmosphere.
Software we used:
- Pycharm
- Panoply (used for view nc files)
- Snap (used for plotting the InSAR data)
Tools we used:
- Tensorflow & Keras
- Jupyter Notebook & Colab-Notebook
- netCDF4
- API (Visual Crossing)
Coding Language:
- Python3 programming language
Space Agency Data
- From Canadian Space Agency we go to Synthetic Aperture Radar Tutorial after we went over the key technologies used in this field, we use as mentioned above a classification model using (VGG-16 architecture) CNN so we can classify the input images (InSAR images) to four categories we mentioned above.
- And also from Synthetic Aperture Radar Tutorial we used an Auto-Encoder to remove the noise due to the water vapor deformation and the key technologies behind that the idea of Transformation Matrix which the auto-encoder already learned it as its learning parameters. An Auto-Encoder function -> f( image + noise) = image.
- We use NASA resources to using netcdf files for San Andreas Fault to doing signal processing using its amplitude and some variables.
Hackathon Journey
Our journey with NASA Space Apps was very enjoyable and fruitful. We learned a lot about INSAR: what it is, its importance, the problems facing earth scientists in it, the ways to solve these problems, and the implementation of these methods using algorithms and ML
We chose this challenge because we enjoy Earth sciences and some of us specialize in it, so we wanted to benefit more through our participation and benefit the world by offering solutions to this challenge.
We've used some weather evidence to confirm the proportion of water vapor in the upper atmosphere first and then determine whether or not to make a correction of the data. The debugging process starts with a classification of data using the ML model followed by the signal processing followed by the Auto-Encoder ML model
Our team had a spirit of teamwork and we were looking for solutions with each other to the problems facing us in parallel, and we were benefiting from the guidance of our observers.
I would like to thank all NASA employees for their creative and innovative young support to create genius solutions for space and Earth.
References
- Shallow creep on San Andreas fault: multiple dates, California, USA from NASA Resources
- Deep Learning Patterns and Practices
- https://weather.visualcrossing.com/VisualCrossingWebServices/rest/services
- https://github.com/ashushekar/VGG16
Tags
#software #intermediate #advanced #earth #deeplearning #classification #machinelearning #signalprocessing