1. Why this trajectory matters¶
Every paradigm shift in this field lowered the amount of domain expertise and labeled data required to build something useful. Knowing where a method sits on that trajectory tells you, roughly, how much labeling work you are signing up for.
2. Core idea¶
The field moved from hand-crafted features, to features learned end-to-end from raw imagery, to general-purpose representations that transfer across tasks with comparatively little fine-tuning. Each shift traded some interpretability and control for less labeling effort.
3. The evolution of paradigms¶
A. Traditional machine learning¶
Random forests, support vector machines (SVM), and gradient boosting dominated through the 2000s and early 2010s. They require hand-crafted features (spectral indices, texture measures) and work well on tabular or small-scale image data. They remain a reasonable choice when labeled data is scarce, compute is limited, or interpretability matters, since a random forest’s feature importances are much easier to explain in a project report than a neural network’s internal weights.
B. Convolutional neural networks¶
Starting around 2015, architectures like U-Net learned features directly from raw imagery, removing the need for manual feature engineering and improving performance on segmentation and detection substantially. The trade-off: CNN-based methods typically need large labeled datasets and task-specific training. Benchmark datasets and the availability of GPUs made this practical at scale.
C. Foundation models¶
Foundation models are large models pre-trained on massive, diverse datasets that learn general-purpose representations. SAM, trained on roughly a billion masks, can segment nearly any object without task-specific training. Geospatial-specific foundation models exist too, pre-trained on large volumes of satellite imagery and adaptable to tasks from flood mapping to crop classification with far less labeled data than training from scratch would need. <!-- TODO: add current course-approved examples and links once a stable list of geospatial foundation models is confirmed for this term -->
D. Vision-language models¶
Models such as CLIP connect visual understanding with natural language, matching images to text descriptions. This enables zero-shot classification: finding images that match a plain-language description (“show me deforestation”) without any labeled training examples for that specific query. This lowers the barrier further, since you no longer need to curate a labeled dataset before you can start exploring a question.
E. AI agents¶
The newest, still-emerging layer combines multiple models and tools with reasoning steps to carry out a multi-step workflow autonomously. An AI agent given a request like “map buildings damaged by the earthquake in this region” could, in principle, chain together imagery retrieval, change detection, building segmentation, and damage classification without a person manually running each step.
4. Python reactivation¶
No code here either. The relevant reactivation is conceptual: think of “fine-tuning” the way you already think of reusing a function with different arguments, you are not rewriting the whole model, only adapting part of it to your data. You will see this pattern concretely with fine-tuning and transfer learning in later lessons.
5. Common pitfalls¶
Assuming a foundation model will work well on your data with zero adaptation. Zero-shot and lightly fine-tuned performance both still depend on how similar your imagery is to what the model saw during pre-training.
Choosing a foundation model purely because it is the newest option. If your labeled dataset is already large and your task is narrow, a CNN trained specifically for it can still outperform a general-purpose model.
Treating interpretability as unimportant. If your project’s argument depends on explaining why the model predicted something, a traditional method with clear feature importances may serve your report better than a large opaque model.
6. Mini task¶
Imagine a project with very little labeled data: perhaps 40 hand-labeled examples. Would you start with traditional machine learning and hand-crafted features, a CNN trained from scratch, or a fine-tuned foundation model? Justify your choice in two to three sentences, considering labeled-data volume, compute, and interpretability.
Sample solution
With only 40 labeled examples, training a CNN from scratch is unlikely to work well; CNNs generally need far more labeled data to learn useful features on their own. A fine-tuned foundation model is the stronger starting point, since it already carries general-purpose representations learned from a much larger pre-training dataset and typically needs only a small amount of task-specific data to adapt. Traditional machine learning with hand-crafted features remains a reasonable fallback if interpretability is a priority for the final report, or if a fine-tuned foundation model is not available for the specific sensor or region involved.
7. Key takeaways¶
The field moved from hand-crafted features, to CNNs that learn features directly, to foundation models that transfer across tasks, to emerging AI agents that orchestrate multi-step workflows.
Each shift reduced the labeled-data requirement, at some cost to interpretability and control.
Traditional machine learning remains a valid choice when data is scarce, compute is limited, or interpretability is a priority.
Matching a method’s paradigm to your project’s realistic data volume and timeline is a project-scoping decision, not just a technical preference.