This course exposes the students to the physical principles underlying satellite observations of Earth by passive sensors, as well as parallel Deep Learning (DL) algorithms that scale on High Performance Computing (HPC) systems.
For the different theoretical concepts (represented by 4 modules), the course provides hands-on exercises. These exercises are part of a project in the context of Remote Sensing (RS) image classification that the students are asked to develop during the whole duration of the course (see the official page of the course).
Upon completion of the course, the student should:
- Know the fundamental laws of Remote Sensing (RS)
- Be able to define properties of satellite images
- Be capable to convert raw satellite images to a format suitable for Machine Learning
- Know how to enable distributed Deep Learning in HPC
- Be able to produce and evaluate land-cover classification maps
Module 1: Introduction to Remote Sensing Systems and Data
This module focuses on the field of RS and Earth Observation (EO), which is primarily concerned with obtaining information about the surface of the Earth from a distance (i.e., from a satellite platform). This module explains what a RS system is and describes its individual processing steps (from acquisition to data generation).
Lecture: Remote Sensing Systems and Data
This lecture focuses on the field of Remote Sensing (RS) and Earth Observation (EO), which is primarily concerned with obtaining information about the surface of the Earth from a distance (i.e., from a satellite platform). In general, it is possible to make a distinction between two very essential ways to acquire RS data. To generate information about a land surface, it is possible to use either Electromagnetic (EM) radiation that originates from the sun (passive RS) or exploit machinery to send out pulses, which return characteristics that can be analyzed (active RS). This lecture introduces the physical principles underlying satellite observations by passive RS and explains its individual processing steps from acquisition to data generation. Furthermore, as a RS instrument comes with a variety of specifications that define what kind of data is gathered and consequently what it can be used for, this lecture will also introduce the differences and interdependencies between the four resolutions: spatial, temporal, spectral and radiometric.
- Theory and fundamental laws of Remote Sensing (RS)
- Introduction to the electromagnetic spectrum
- Interaction of EM radiation with atmosphere and Earth's surface
- Properties and resolutions of RS data
Laboratory: Copernicus Sentinel-2 Mission
This laboratory first introduces the Copernicus Sentinel-2 mission and its properties. Then it shows the necessary steps to find and download Sentinel 2 satellite images from the Copernicus Open Access Hub. It will demonstrate how to define a search area, add search filters, and then how to observe the metadata to finally download the images. It will finally show how to use QGIS to visualize the retrieved images.
- Introduction to Copernicus Sentinel-2 Mission
- Properties of Sentinel-2 Data
- Data Retrieval with the Copernicus Open Access Hub
- Data visualization in QGIS
Laboratory: Copernicus Sentinel-2 Mission
This laboratory introduces the usage of APIs to retrieve acquisitions of Sentinel-2 for specific geographical locations and dates. The student will become familiar with the metadata used for data retrieval.
- Overview of Jupyter-JSC
- Learning the usage of a Python API for the retrieval of Sentinel data (f.e. Sentinelsat)
Module 2: Remote Sensing Data Processing
This module introduces open-source RS libraries that are indispensable to work with Sentinel-2 satellite images. To cope with the large sizes of the data, the students are granted access to Jupyter-JSC (i.e., a web-based environment for Jupyter notebooks that is connected to the HPC systems of the Jülich Supercomputing Centre).
Lecture: Machine Learning for Classification
This lecture introduces the fundamentals at the basis of end-to-end machine learning in the context of remote sensing data classification.
- Benefits and limitations of machine learning in remote sensing applications
- Importance of the extraction of informative features
- The components of the learning process
Module 3: Distributed Deep Learning for Classification
This module covers DL methods and how to unlock computing resources of supercomputers to speed-up training and inference processes. The students work through an end-to-end DL model. They train and optimize the model for a use case on land cover classification with time series of Sentinel-2 images.
- Deploy a DL Model for classification tasks
- HPC system architectures
- Distributed DL with HPC
Module 4: Performance Measurement and Metrics
This module presents the necessary tools to evaluate the results. It starts by introducing standard metrics for classification derived from the confusion matrix (Overall Accuracy, k coefficient, F-Score) and concludes with more technical evaluation procedures to verify the spatial and temporal consistency of the classification maps.
- Evaluation with metrics
- Summarizing results in a concise manner
- Assessing the quality of classification maps