Thursday from 8:30AM to 11:30PM
- Instructor: Guray Erus, [email protected]
- Teaching Assistant: Tristan Grupp, [email protected]
- Instructor: Friday 3PM to 5PM on Zoom
https://pennmedicine.zoom.us/j/92234699148
Meeting ID: 92234699148
- GitHub: https://github.com/MUSA-650/MUSA-650_Spring2024
- Canvas: https://canvas.upenn.edu/courses/1770699
Satellite remote sensing is the science of converting raw aerial imagery into actionable intelligence about the built and natural environment. This course will provide students the foundation necessary for application of machine learning algorithms on satellite imagery. Use cases include building footprint detection, multi-class object detection in cities and land cover/land use classification. The students will learn basic concepts of machine learning, including unsupervised and supervised learning, model selection, feature elimination, cross-validation and performance evaluation. After learning traditional methods and algorithms, the course will focus on recent deep learning methods using convolutional neural networks and their application on semantic image segmentation.
Students are expected to be familiar with and comply with Penn’s Code of Academic Integrity, which is available online at https://catalog.upenn.edu/pennbook/code-of-academic-integrity . Any form of academic dishonesty will be penalized with a failing grade for the assignment in which the infringement occurred. Additionally, any violations of the Code will be referred to the Office of Student Conduct for further disciplinary action.
The main learning goal of this class is to provide students essential knowledge and skills in applied machine learning. This will be done through hands-on applications and a large number of practical examples. While the class will cover general concepts of machine learning, applications will specifically focus on use cases in remote sensing. There will be a specific focus on deep learning applications, as deep learning is becoming an essential framework in machine learning, with state-of-the-art performance. The students will learn how to define a problem, select appropriate algorithms and tools, design and implement their machine learning models, and apply and validate their models on new datasets. This is a hands-on class involving multiple examples that will be explained and run real-time during the lectures. Active participation on Google Colab is highly recommended both during and out of lecture times.
The grading breakdown is as follows: 40% for homework (4HW), 30% for project (1 PRJ), 20% for participation and %10 for final quiz. Homework and project grades are not directly linked to accuracy of the final applications. Multiple factors, including data organization, model selection and presentation of the results will be evaluated.
Homework assignments will involve implementation of machine learning applications using toy datasets provided by the instructor. Homework is assigned before the end of a class and is due at the end of the following week's class (unless indicated otherwise). Late homework will be accepted but penalized. Students are encouraged to work in groups, but they must submit a homework assignment that is uniquely theirs.
For the term project, students are expected to make their own research to select a machine learning problem in remote sensing, collect and organize their datasets, and implement their solution, working in groups or individually.
Students are encouraged to ask and answer questions. Participation grade is a function of both in-class and out-of-class participation.
This course relies on use of Python and various related packages. The class will include interactive demonstrations on Google Colab. All software is open-source and freely available.
| Class # | Date | Topic | HW & Project |
|---|---|---|---|
| Week 1 | 1/18 | Overview of remote sensing and satellite imagery appications | |
| Week 2 | 1/25 | Fundamentals of machine learning from a remote sensing perspective | |
| Week 3 | 2/1 | Data preparation: imaging feature extraction, visualization, normalization, data harmonization | HW1 |
| Week 4 | 2/8 | Dimensionality reduction and unsupervised learning | |
| Week 5 | 2/15 | Supervised learning for land cover classification, Part 1: training, cross-validation, method and model selection, parameter optimization | HW2 |
| Week 6 | 2/22 | Supervised learning for land cover classification, Part 2: validation, evaluation, multi-class classification | |
| Week 7 | 2/29 | Ensemble methods in machine learning. Case studies: DSTL and EuroSat challenges | HW3 |
| Spring Break | |||
| Week 8 | 3/14 | Fundamentals of deep learning | |
| Week 9 | 3/21 | Recent advances in deep learning: literature review and paper discussion | HW4 |
| Week 10 | 3/28 | Convolutional neural networks for image classification; UNet architecture for semantic segmentation | Project Proposals |
| Week 11 | 4/4 | Case studies: DSTL and EuroSat challenges revisited | |
| Week 12 | 4/11 | Case study: predictive modeling using deep learning | |
| Week 13 | 4/18 | Big data and machine learning: techniques, tools, challenges, future directions | |
| Week 14 | 4/25 | Reviews | Project Presentations |