CASCADE

CASCADE (An Onboard Crop Anomaly Screening, Confirmation, and Alert Downlink Engine) turns crop-stress monitoring into an onboard scheduling problem: a Bayesian belief gate drives a finite-horizon MDP over four actions (SKIP, MOD13, FUSE, FUSE_PRIORITY) while fusing MOD13A1 EVI, MOD11A1 LST, and MOD09A1-derived NDWI into a three-channel Crop Stress Composite (CSC).

Highlights

• Paper-aligned headline metrics (100-seed Monte Carlo): 99.1% downlink reduction, 38.3% energy saving, 0.6% FP rate, 92.5% / 49.2% CPU (peak / seasonal).
• Real-scene replay anchors (Westlands/Firebaugh, CA):
  • 2024 quiet season: peak CSC 0.412 vs threshold 0.615 → no FUSE_PRIORITY expected.
  • 2014 D4 drought: peak CSC 0.869, FUSE_PRIORITY at 6/6 windows → 2.1× the 2024 peak (no retuning).
• One-command reproducibility: make repro-2014, make repro-2024, make test.
• Canonical claims protected by tests: failing a paper anchor fails CI.

Watch the CASCADE video on Vimeo

Overview

CASCADE is a NASA Space to Soil submission repository containing:

• Synthetic benchmark: a 100-seed Monte Carlo study (ROC sweep + ablations + sensitivity).
• Real-scene MODIS replay: AppEEARS-backed replays over the Westlands AOI.
• CSC calibration: offline parameter search with accepted calibration fixtures.
• Economics model: a rollout model and accepted baseline fixtures.
• Paper source: paper/EmilLambert_CASCADE.tex and the compiled PDF.

Verified headline metrics (paper numbers)

vs raw downlink	vs fixed onboard	Recall	FP rate	CPU (peak / seasonal)
99.1% downlink reduction 38.3% energy saving	77.6% downlink reduction 25.4% energy saving	100.0%	0.6%	92.5% / 49.2%

Installation

Reviewer target: Python >=3.10,<3.13 (fresh virtual environment).

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt -r requirements-dev.txt

The source-tree shim supports python -m cascade.* directly from a checkout. If you want an editable package install as well, run:

python -m pip install -e .

Usage (fast path)

Run the synthetic benchmark (skips slow “additional ablations” by default):

python -m cascade.simulate

Run the paper-anchor replays. These prefer tracked artifacts when present, so they work offline and return the same JSON payloads used by tests:

python -m cascade.replay --year 2014
python -m cascade.replay --year 2024

Run tests:

python -m pytest -q

Run only the accepted-artifact validation checks:

python -m pytest -m validation

Makefile (judge-friendly)

make test
make repro-2014
make repro-2024

After reviewing regenerated files under build/, promote selected outputs into tracked artifacts/ and paper/figures/:

make figures

The Makefile uses .venv/bin/python when available and otherwise falls back to python3; pass PYTHON=/path/to/python to override it.

Real MODIS replay (Earthdata/AppEEARS)

For live downloads, set Earthdata credentials in .env or your shell:

export EARTHDATA_USERNAME='your-username'
read -s EARTHDATA_PASSWORD
export EARTHDATA_PASSWORD

Test AppEEARS login:

curl -i -u "$EARTHDATA_USERNAME:$EARTHDATA_PASSWORD" \
  -X POST https://appeears.earthdatacloud.nasa.gov/api/login

Run the lower-level live replay workflow directly:

python real_modis_replay.py --help

The offline reviewer replay commands above are preferred for quick evaluation; live replay writes fresh outputs to build/replay/ and may reuse data/cache/.

Additional workflows

Regenerate CSC calibration outputs:

python scripts/calibrate_csc.py

Regenerate unit-economics outputs:

python scripts/unit_economics.py

Repository map

• Package code: src/cascade/
• Replay: src/cascade/replay/modis.py
• Benchmark: src/cascade/simulation.py
• Calibration: src/cascade/calibration.py
• Economics: src/cascade/economics.py
• Tracked outputs: artifacts/
• Transient outputs: build/
• Paper: paper/EmilLambert_CASCADE.tex (source of truth)

More detail: docs/repo_structure.md

Concise docs

• Repository structure
• Reproducibility
• Validation
• Replay anchors
• Unit economics

For judges — 5-minute evaluation path

1. Read Highlights + the metrics table above.
2. Open artifacts/benchmark/roc.png.
3. Run make repro-2014 or make repro-2024 (or python -m cascade.simulate).
4. For artifact regression checks, run python -m pytest -m validation.
5. Open the tracked replay metrics under artifacts/replay/ or read docs/replay_anchors.md.

Citation

If you use this repository, please cite it via CITATION.cff.

License

Released under the MIT License. See LICENSE.