Machine Learning for Processing Satellite Images

(joint work with Marian Neagul and Gabriel Iuhasz)

Semantic segmentation plays an important role in the area of satellite image processing. In this presentation, we analyze different network topologies based on convolutional neural networks. Most of the existing solutions, do not tackle the problems related to image preprocessing, focusing instead on the topology of the network and its hyper-parameters. we provide a suite of tools tailored for Earth Observation data, tools that aim to provide the required services for supporting easy experimentation and integration of various Deep Learning models, preprocessing techniques and model ensemble methods. We employ our tools and state of the art machine learning models for large scale image segmentation tasks like building footprint detection. In our study we focus on extending and evaluating state of the art deep learning models (like U-Net, Segnet) for Earth Observation tasks. We integrate Machine Learning and Computer Vision tools like TensorFlow, Keras, OpenCV and SciKit-Learn with modern Earth Observation tools like RasterIO (for raster and elevation model handling).

Also, in this presentation, we are going to show how to use libraries like RasterIO, OpenCV, Pillow, OTB, scikit-image, Keras integrated with Jupyter. Also, we are going to discuss how one can view satellite images in Jupyter .