Sugarcane next-year forecast (v2) — Louisiana
Predicts whether a cropland pixel will be sugarcane next year from this year's and last year's Google Satellite Embedding vectors (2 × 64 dims) plus this year's and last year's cane/not-cane status. Gradient-boosted trees (sklearn HistGradientBoosting). Labels: USDA Cropland Data Layer.
Why this is non-trivial
Louisiana cane rotates fast: 24% of cropland pixels change cane status each year, so "same as last year" (persistence) is only 76–78% accurate and is wrong on every change by construction.
Results (held out by year; train on earlier transitions only)
| Test transition | Accuracy | Persistence | New-cane recall | New-cane precision | Removal recall | Removal precision |
|---|---|---|---|---|---|---|
| 2022 → 2023 | 0.881 | 0.780 | — | 0.884 | — | 0.766 |
| 2023 → 2024 | 0.857 | 0.760 | 0.946 | 0.832 | 0.253 | 0.755 |
Expansion is predictable, contraction isn't. Fields coming into cane (after fallow/soy rotation) are visible a year ahead; fields leaving cane are a post-harvest decision with little prior signature.
Scope & caveats
- Louisiana only, 30k points stratified 50/50 on 2021 cane status, 2018–2024 (5 training transitions).
- Labels are CDL (map-derived); some "changes" are label noise.
- Not validated outside Louisiana. Embeddings are region-specific (see the v1 classifier card), so retrain with local labels before using elsewhere.
Data
data/train.parquet: one row per point with A00_2018 … A63_2024, label_2018 … label_2024,
code_YYYY (CDL class), lon, lat. Feature order for the model is in metrics.json.
Embeddings: Google Satellite Embedding V1 (CC-BY-4.0). Labels: USDA NASS CDL (public domain).
- Downloads last month
- -