satellite-image-deep-learning/techniques

This repository is a reference collection for people who want to apply deep learning to satellite and aerial images. It does not contain runnable code of its own. Instead, it links out to dozens of external projects, papers, and tutorials, organized by task type. The intended audience is researchers, engineers, and students working in earth observation or geospatial analysis who already have some familiarity with machine learning concepts. The topics covered span a wide range of problems that come up when working with imagery from satellites or aircraft. Classification means assigning a category label to a whole image or a region, such as identifying whether a patch of land is urban, forest, or agricultural. Segmentation goes further by labeling each individual pixel. Object detection means finding and drawing boxes around specific things in an image, like buildings, vehicles, or ships. Beyond those core tasks, the repository also covers regression (predicting a continuous value from imagery), cloud detection and removal, change detection between images taken at different times, time series analysis, and crop classification and yield forecasting. There are also sections on more specialized techniques. Generative networks can synthesize new satellite images or translate between different sensor types. Autoencoders and similar methods compress image data into compact representations useful for similarity search. Few-shot and zero-shot learning approaches address the common problem of having very little labeled training data for a specific task. Self-supervised and contrastive learning methods let models train on large amounts of unlabeled imagery. SAR (synthetic aperture radar) is a radar-based imaging technology distinct from optical cameras, and the repository includes links specific to it. There are also sections on explainability tools, large vision and language models applied to satellite data, and foundational models built for geospatial use. Each section is a curated list of links to GitHub repositories, academic papers, and blog posts. The README is designed to be searched rather than read top to bottom. The maintainers suggest using browser search to jump to the topic you need. The full README is longer than what was shown.