Does anyone know about applications/demos of OpenVINO to classify satellite imagery?
I'm interested to explore this tool in this field. By watching some of the demos on using it I wonder if by training a model using videos of a field expert while navigating in Google Earth could be used as data input.
Clips out of Google Earth could be used to inform/train the model over a specific type of classification problem.
Does this sound like a potential application of OpenVINO?
Thanks for reaching out to us.
As of now, we don’t have a specific demo that classifies satellite imagery. However, your idea to train a model using videos while navigating in Google Earth as data input is enticing, and we would encourage you to pursue it.
You can use an optimized framework (TensorFlow, Caffe, MXNet, etc.). I would suggest you find a related Neural Network model and train it. You can check out Intel and Public pre-trained models as well.
Your trained model can then be optimized using Model Optimizer, which converts the model to an Intermediate Representation (IR), which is represented in a pair of files (.xml and .bin). These files describe the network topology and contain the weights and biases binary data of the model.
You can then use Inference Engine to run inference and output results on multiple processors, accelerators, and environments with a write once, deploy anywhere efficiency.
Refer to the following page, Integrate the Inference Engine with Your Application, which explains the process of integrating the Inference Engine into your application.
Do have a look at the following article, which discusses approaches that demonstrate the feasibility of deep learning in automated satellite image annotation.
Please also refer to the following three papers, which make use of available satellite data in machine learning:
Additionally, check out the following article on the first-ever hardware-accelerated AI inference of Earth observation images on an in-orbit satellite.
And do browse through the AI Solutions Catalog, where you can find optimized and market-ready partner solutions to support your AI deployment needs.
Thank you, Munesh.
Thank you for the references. Plenty of things to chew on.
I've seen several applications of deep-learning using satellite imagery. However, one thing that I've seen is that in most cases these classifications or segmentations are binary (e.g. roofs | no-roofs, roads | no-roads, etc).
In this case, I describe, is more a group of features observed by the expert that support the decision for the final classification.
A common task from a field expert on our side would-be, taking an IPad with Google Earth Pro on it and walk around an area and do some field observations around his actual location, then start drawing a polygon and set the classification to a specific polygon(boundary) (not really on a pixel basis) based on different features like the type of vegetation in the area, vegetation density, amount of bare soil, knowledge of the soil (based on soils database), and other experts' insights.
I learned about OpenVINO through the case at John Deere where they use an algorithm for detecting issues with welding quality (i.e. gas bubbles trapped).
This is when I thought about our needs since the pattern in the resulting welding seems to be a complex one, I guess. The only thing that the resulting model sees, to perform the inference of defective | non-defective welding, is the fed clips (real-time?) of the final welding process. I'm sure this is an over-simplifaction of that process, but in the end, it sounds like a pattern recognition application that I thought could be, in some way, applied to a case like the one I described, or maybe I'm just absolutely wrong.
It could worth trying to set a research project to explore this idea.
Thank you for waiting and apologies for the late reply.
The developers have taken note of the evolving requirements of satellite image classification, and are open to review such contributions in the form of a pull request in the OpenVINO GitHub at https://github.com/openvinotoolkit/openvino/pulls, if it meets our legal requirements.
Thank you for sharing your ideas with OpenVINO Community.