Story County tiles
Each square is a 1.2 km analysis tile built from 30 m USDA CDL pixels. Click a tile, or pick a featured one. The map shows 2025 crop groups.
County typical tile
Your priorities
Move the sliders to say what matters on this farm. Scores update immediately. Equal weights are the default.
Rotation strategies compared
At least three options, including the current corn-soy rotation, plus non-dominated sequences from a small NSGA-II search. Ranked for the selected tile and your weights.
Trade-off table
Pareto front
Each point is a 4-year sequence on the selected tile. X is water demand, Y is erosion risk (lower is better on both axes). Gold points are named strategies. Green points are other non-dominated NSGA-II options.
NASA POWER climate context (county center)
Growing-season (Apr–Sep) precipitation and growing degree days at 42.035°N, 93.465°W, 2000–2024. Planning indicators use the median of the last ten years, not a weather forecast.
What each number means
Cover-crop “+cover” years assume a winter rye that shrinks bare-soil months and uses a little extra water. That assumption is documented, not observed as a CDL class. SMAP, ECOSTRESS, GPM IMERG and GRACE are named in the challenge and are not in this local demo because they need an Earthdata login.
Next-season crop, test years 2024–2025
Persistence is weak because Story County already alternates corn and soy. The practical bar is the 2nd-order Markov table. LightGBM is useful when it improves calibration and rare classes, not just headline accuracy.
Feature-group ablation
Calibration (temperature scaling on 2022–2023)
County maps (local run)
Colab Full-County Run (2024–25)
A larger personal-Colab CPU run on the same Story County AOI and the same temporal split (train through 2021, validation 2022–2023, test 2024–2025). Seed 0, 25 Optuna trials, 400,000 sampled pixels (266,664 / 66,666 / 66,666), plus a 2025 full-county map over about 1.50 million farmed pixels. Shown here as evidence alongside the local Story County run above, not as a replacement for it. Both runs are labeled one seed, not a published benchmark.
Baselines vs LightGBM (Colab test years)
Feature-group ablation (Colab)
Confusion and calibration (Colab)
2025 full-county map (Colab)
Held-out year 2025: USDA CDL actual, LightGBM prediction from history + soil + climate + NDVI, and disagreement (30.2% of farmed pixels). Pixel accuracy over 1,495,561 farmed pixels is 0.811. One seed, not a published benchmark.
Data and software
Original software in this folder is licensed under Apache License 2.0. All sources are cited above. Builds on our earlier code and open-source libraries. One seed, not a published benchmark.