A full year of 15-day optical and radar observations for every parcel in a southern Punjab cane belt, plus field-surveyed sugarcane points from 48 districts of Punjab and Sindh. Built for parcel-level crop and land-use classification where fields are small, mixed, and harvested on very different calendars.

step	what happens
01 · Sources	Register parcels of a southern Punjab cane belt (label year 2025) and field-survey points across Punjab & Sindh (2023).
02 · Imagery	Sentinel-2 optical, masked for cloud, shadow, cirrus and snow; Sentinel-1 radar, speckle-filtered — one full year back from 31 December.
03 · Compositing	24 optical and 24 radar windows of 15 days, median per window, averaged over each parcel. Empty windows stay empty.
04 · Features	14 vegetation indices, 4 radar bands, within-parcel spread, then peak, amplitude and harvest-drop timing.
05 · Table	One row per parcel: 549 columns with the class label, spatial group keys and coordinates for blocked cross-validation.

Sugarcane is the only field crop that stays green for ten to twelve months and then falls off a cliff when it is cut. Orchards never fall; wheat-dominated cropland peaks when cane is at its lowest. That contrast is what the dataset is built to capture.

The seven classes

Class	Rows	Share	What it is
Sugarcane	348,165	43.9%	Standing and ratoon cane; three harvest calendars — winter (92%), late spring (7%), harvested before the anchor (1%)
Agri	240,408	30.3%	All other cropland: wheat, cotton, rice, fodder, rotation fallow
Builtup	106,706	13.4%	Houses, farmsteads, sheds and yards inside the register
Open Land	61,823	7.8%	Barren and uncultivated ground, sand
Orchard	34,215	4.3%	Mango, citrus, guava and other perennial fruit — the hardest confuser for cane
Road	2,080	0.3%	Metalled and unmetalled tracks
Waterbody	346	0.04%	Ponds, channels and river sections

Typical signatures: cane peaks near 0.80 NDVI in Aug–Oct with one deep drop at harvest and about −15 dB VH at canopy peak; orchards hold 0.5–0.8 all year with no harvest drop; wheat-led cropland peaks in Feb–Mar and is bare in Apr–May; built-up sits flat at 0.2–0.4, open land at 0.1–0.2, water below zero with high MNDWI.

Two sources, one schema

source	rows	label year	geometry	labels
`parcels`	785,709	2025	register polygons, as registered — no inward buffer, no minimum size	grower declaration, all 7 classes
`gtp`	8,034	2023	field-survey points buffered to 15 m circles	field observation, sugarcane only

Parcels follow an acre grid, so many contain a bund, a house or a strip of another crop; sugarcane parcels carry a cane_frac of 0.70–1.00 (median 0.95). This is deliberate — it is what production inference sees. The field points span 48 districts from 24.4°N to 32.9°N and carry survey date, season and growth stage; they bring the crop's regional variety — autumn-planted cane, spring cane, ratoon — that a single belt cannot supply.

What each parcel holds

Optical, 24 × 15 days. P00_* is 17–31 December, P23_* is early January. Indices: NDVI, NDRE, LSWI, EVI, SAVI, IRECI, REIP, GNDVI, NBR, NDTI, NDBI, BSI, NDWI, MNDWI, plus Pxx_NDVI_std and Pxx_LSWI_std (within-parcel spread).

Radar, 24 × 15 days. M00_* … M23_*: VV, VH (dB), VV−VH, RVI, plus Mxx_VH_std. One orbit pass per source, 30 m circular speckle median on the composite. Radar sees through cloud, so the harvest drop stays visible through the monsoon and winter fog.

Phenology and coverage. Peak value and its timing, amplitude, area under the curve, number of green periods, the largest drop and rise with their dates — the harvest signal — and how many periods actually had an observation.

Labels and keys. sample_id, source, label_year, v5_class, is_cane, harvest_calendar, cane_frac, parcel_ac, lon, lat, spatial group keys, and for survey rows province, district, season, growth_stage.

Missing observations are kept as missing. A fortnight with no cloud-free image is left empty rather than filled in, so a model can learn the difference between "bare ground" and "nothing was seen".

Honest limits

Labels are declarations. Register entries are grower declarations, not field checks. Two noise pockets are known: about 2,600 "Agri" parcels stay green for twenty periods or more (very likely cane), and about 2,000 "Sugarcane" parcels never pass 0.5 NDVI.
Coverage gaps. Optical gaps are 0.2% for 2025 and 1.6% for 2023 (a wet August). Sixteen extraction batches — 6,400 parcels, 0.8% — are missing on a compute quota and listed in v5_assembly_report.json.
Survey points are positives only. There are no verified non-cane field points in this release, so precision outside the belt cannot be measured from it.
Small parcels. A quarter-acre parcel is two to six Sentinel pixels, most of them on a boundary; expect these to behave differently from half-acre and larger fields, and report them separately.
Split it spatially. Neighbouring parcels share soil, water and planting date. Published results use 5-fold cross-validation on 5 km blocks with a 1 km buffer removed, plus a whole-zone hold-out and a survey-source hold-out. Random splits flatter a model badly.

Files

file	contents
`v5_train_table.parquet`	the table: 793,743 rows × 549 columns, ready for training
`batches/`	2,066 raw extraction batches — the resumable cache the table is built from
`samples/*.parquet`	per-sample attributes without geometry
`samples/*.gpkg`	geometries, EPSG:4326
`v5_2_oof_predictions.parquet`	out-of-fold sugarcane probability for every row, from a model that never saw its 5 km block

No personal data: grower names, identity numbers, phone numbers and addresses are excluded from every file.

Trained models: AdilMunawar/XGboost — see spectral_v5_2.

Provenance

Imagery: Copernicus Sentinel-2 L2A and Sentinel-1 GRD via Google Earth Engine (COPERNICUS/S2_SR_HARMONIZED, COPERNICUS/S1_GRD). Parcels: grower registers, 2025. Ground truth: land-use field survey, 2023.

Downloads last month: 138

Total file size:

82.3 GB

You need to agree to share your contact information to access this dataset

Sugarcane & land-use parcel dataset — Punjab & Sindh