--- language: - en license: other license_name: upstream-product-terms license_link: https://huggingface.co/datasets/PHENOBIID/CropDynamicsBench/blob/main/TERMS.md tags: - agriculture - crop-yield - earth-observation - time-series - weather pretty_name: CropDynamicsBench --- # CropDynamicsBench Processed global data for weather-conditioned vegetation forecasting and in-season yield assessment of maize, rice, soybean, and wheat. NDVI and GPP span 1982--2016; yield history, ERA5-Land, and legacy LAI include 1981--2016. This is the numeric data release, not a collection of model outputs. ## Download Files are losslessly compressed tar shards. Their members preserve the `Data/` layout used by the processing code. `manifest.json` lists every member's shape, dtype, size and SHA-256, as well as the checksum of each archive. The release contains **2,691 files in 52 shards**: **5.48 GiB compressed** and **44.40 GiB extracted**, including the optional SEAS5 supplement. Reserve about 60 GiB for the download cache and extracted data together. Select product prefixes in `shards/` or use the companion code repository's `scripts/download_processed_data.py`. All shards can also be obtained with: ```python from huggingface_hub import snapshot_download snapshot_download(repo_id="PHENOBIID/CropDynamicsBench", repo_type="dataset", local_dir="CropDynamicsBench_download") ``` The download contains archives, not automatically extracted NumPy arrays. Use the verified extraction script from the code release to restore `Data/`. For exact reproduction, pin a dataset commit rather than the moving `main` branch. ## Contents and shapes | Shard prefix | Contents | Typical shape | |---|---|---| | `crop_active` | Four-crop yield, aligned ERA5-Land, MIRCA-OS month mappings and areas | weather `[13,12,360,720]`; yield `[360,720]` | | `era5_land` | Monthly 0.5-degree weather before crop-active packing | `[12,360,720]` per variable/year | | `gdhy` | Four-crop annual yield with reordered longitudes | `[360,720]` | | `ndvi` | Monthly and crop-active NDVI, availability and quality fractions | `[12,360,720]` | | `gpp` | Monthly and crop-active GPP totals, daily rates and valid-area fractions | `[12,360,720]` | | `lai` | Legacy monthly and crop-active LAI used for cohort support and comparisons | `[12,360,720]` | | `seas5` | Optional ensemble-mean seasonal weather on its native 1-degree grid | `[6,180,360,3]` per initialization | The 0.5-degree latitude and longitude arrays are shipped with the products: latitude increases from -89.75 to 89.75, longitude from -179.75 to 179.75. Do not replace crop-active slots by calendar-month indices. `src_rel.npy`, `month_rel.npy`, `valid_rel.npy`, and `weight_rel.npy` define the packed support; invalid slots and missing values must be masked before model training. The mapping arrays have shape `[12,360,720]`: `month_rel` stores calendar months 1--12 (0 for padding), `src_rel` stores zero-based source indices 0--11 (255 for padding), and `valid_rel` is 1 for active slots. Active months are packed in ascending calendar-month order. `weight_rel` is monthly growing area divided by the sum of growing areas over valid months at that grid cell. NDVI is dimensionless, yield is t/ha, and the selected GPP state is the daily mean rate in gC m-2 day-1, not the monthly total. ERA5-Land channel order is also supplied as `variable_order.txt`: `d2m, t2m, stl1, stl2, swvl1, swvl2, swvl3, ssrd, pev, u10, v10, sp, tp`. These are physical processed arrays; model normalization must be fitted on the corresponding training partition, not all years. ### Optional seasonal forecasts SEAS5 contains 432 monthly initializations (1981--2016), six lead months and 25-member ensemble means. Channels are `t2m, tp, ssrd`, in K, m/day, and J/m2/day. Latitude descends from 89.5 to -89.5; longitude increases from 0.5 to 359.5. These files are **native-grid forecasts**, not already crop-aligned or bias-corrected model inputs. Use the separate SEAS5 crop-weather interface and training-only calibration before comparisons. The main benchmark uses realized ERA5-Land weather and is a retrospective conditional assessment. ## Evaluation protocol The companion `protocol.json` defines the temporal splits. Model fitting through 2001, 2005 and 2009 is paired with evaluation in 2002--2004, 2006--2008 and 2010--2016, respectively. The three blocks comprise 13 evaluation years. Seeds are 42, 45 and 48; primary nominal hidden fractions are 10%, 30%, 50% and 70%. Future hidden vegetation is a target, not a permitted model input. ## Scope All selected numeric arrays are preserved byte-for-byte. Duplicate smoke-test directories, intermediate caches, private machine-specific audit logs, model weights, papers and raw-provider archives are excluded. Generated release checksums replace private audit paths, not the original processing history. The full numeric release does not by itself supply every legacy training cache or establish a fresh full-benchmark retraining reproduction. The companion repository documents the verified sample inference and reconstruction stages. ## Sources and terms See `sources.json` for original provider records, versions, and attribution, and [TERMS.md](TERMS.md) before reuse. This release does not place all upstream products under a new blanket license. In particular, rEC-LUE GPP is a model-derived productivity product, not an independent direct measurement.