satellite-image-deep-learning/techniques: a curated index, not a library

What the repository actually is

This is a catalogue. The README opens with a link to satellite-image-deep-learning.com and then presents a table of contents of task categories: classification, segmentation, object detection, regression, cloud detection and removal, change detection, time series, crop classification, crop yield and vegetation forecasting, generative networks, autoencoders and image embeddings, few and zero shot learning, self-supervised and contrastive learning, SAR, explainable AI, large vision and language models, and foundational models. Under each category the README lists third-party GitHub repositories with a one-line description of what each does.

The audience is anyone starting a remote sensing project who needs to know what already exists before writing a model. That includes engineers moving into Earth observation from a general computer vision background, and researchers looking for a baseline for a specific task. The repository's own instruction for using it is a browser search: use Command + F on Mac or CTRL + F on Windows to search the page for a term such as 'SAM'. That single sentence tells you the intended scale of the document and the intended reading mode.

How the index is structured and how entries are described

The data flow is flat and manual. A maintainer writes a category heading, then adds a bullet with a link and an arrow pointing to a short description. For example, the classification section lists EuroSat-Satellite-CNN-and-ResNet with the note that it classifies custom image datasets by building convolutional and residual networks from scratch in PyTorch, and lists WaterNet as a CNN that identifies water in satellite images. Some entries carry a second link to an accompanying article, as with the Land-Cover-Classification-using-Sentinel-2-Dataset entry, which points to a Medium write-up.

The repository tree matches that: .github/, .gitignore, .mlc_config.json, AGENTS.md, LICENSE, README.md, images/. The .mlc_config.json file is configuration for a markdown link checker, which is the mechanism that keeps the links from rotting, and it explains the v1.3 release note, 'Prune out of date links'. The images/ directory holds illustrations used inside the README, such as the UC Merced dataset sample shown in the classification section. There is no package, no module, no test suite and no model weights anywhere in the tree.

Installing it is not the point, reading it is

There is nothing to install. The README gives no pip command, no conda environment, no Docker image and no setup script, and the repository layout confirms there is no code to run. The README does give one concrete instruction for working with the text itself: use Command + F on Mac or CTRL + F on Windows to search the page for a term such as 'SAM'. That is the only usage step the project documents.

If you want the README available offline so you can search it without a browser, the repository is a plain git checkout of README.md, LICENSE, .mlc_config.json, AGENTS.md and images/. Clone it, then search the file with the same term you would have typed into the browser. After that, open the linked repository for the entry that matches your task, because this repository carries no assessment of build quality, accuracy or maintenance for any of the links.

The limitation: a link is not a recommendation

The README does not state that any linked project is maintained, correct or reproducible. It gives a description, not an evaluation, and descriptions can outlive the code they describe. The v1.3 release was titled 'Prune out of date links', which is evidence that link rot is a recurring problem here rather than a one-off. A category can therefore look well populated while several of its entries point at abandoned work.

There is also a scope mismatch between the framing and the contents. The introduction talks about architectures, models and algorithms, but the artefacts are hyperlinks. If you need a working classifier today, this repository will not give you one, and the time you spend triaging links is time not spent training. The other failure mode is staleness of framing: the category list includes foundational models and large vision and language models, which are recent additions, while older sections retain entries from earlier waves of work. Nothing in the README marks which entries are current and which are historical.

Where a task-specific library beats an index

If your goal is to load Sentinel-2 imagery and train a segmentation model, a framework with a data pipeline is a different kind of tool. TorchGeo is the clearest contrast: it is a library that provides datasets, samplers and transforms for geospatial data, so you import it and run it. This repository instead points you at other people's repositories, and you assemble the pipeline yourself from whichever link looks closest to your problem.

The difference is in what you get on day one. With a library you get an API, versioned releases and issue tracking against the code you are running. With this index you get a starting point and a reading list, plus the freedom to pick a model that matches your sensor and resolution rather than whatever the library happened to implement. For SAR, hyperspectral or crop yield work, that freedom matters, because no single library covers all of those well.

Maintenance, licence and the cost of following links

The last push to the default branch was on 2026-09-04, and the most recent tagged release is v1.3 from 2025-07-05, described as 'Prune out of date links'. Earlier releases are 28.1.2024 ('General tidy up') and 22.6.2023 ('Adds sponsor'). The cadence is irregular and release notes are terse, so the useful signal for a reader is the commit activity on the README rather than the tags.

The repository is licensed Apache-2.0, and the LICENSE file sits at the top level. That covers this repository's own contents, which are text and images. It does not extend to any linked project, each of which carries its own licence, and the README does not summarise those licences. If you plan to use a linked model or dataset commercially, check that project's licence and the dataset's terms separately. The upgrade cost here is low in the mechanical sense, since updating is a git pull, and high in the evaluative sense, since each pull can add or remove entries you were relying on.

Editorial conclusion

Adopt it as a discovery layer when you need to find prior art for a remote sensing task, and treat every link as an unvetted candidate that you must check for maintenance and licence before use. Do not adopt it if you want a runnable pipeline, because the repository contains no model code, no training script and no dataset downloader. Before relying on it, verify what the current README actually lists under your task, and check the licence of each linked project separately, since the Apache-2.0 notice here covers only this repository's own contents.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. README
4. Releases
5. satellite-image-deep-learning/techniques on GitHub