A framework for studying sampling strategies (random, cluster, convenience, cost-aware, active learning) for remote-sensing regression tasks across three datasets: USAVars, India SECC, and Togo soil fertility.
Data download and featurization instructions are in datasets/README.md.
datasets/raw/mosaiksis adapted from Global Policy Lab/mosaiks-paper (feature extraction and dataset handling).sampling/is adapted from TypiClust's deep-al module (active learning and sampling methods) — see sampling/README.md.
- Data - download and featurize each dataset.
- Groups - build geodataframes and group assignments (admin regions, image clusters, NLCD land cover) used by group-aware sampling strategies.
- Initial sample - construct a starting labeled set (random, cluster, or convenience sampling).
- Sampling - train a model while adding to the labeled set with a chosen sampling method.
- Summarize - parse logs and generate tables/figures.