Synthetic soil spectral dataset for predicting nutrient content from reflectance data. I'm based in Austria and got interested in precision agriculture after talking to some farmers in Lower Austria who were paying a fortune for lab-based soil tests. The idea here was to create a training dataset that captures the relationship between what a spectrometer "sees" and what's actually in the soil — nitrogen, phosphorus, potassium, organic matter, pH.

The dataset covers 8 soil types common in Central Europe, with a focus on Austrian regions. The physics is simplified but the spectral shapes are realistic — darker soils have lower reflectance, organic matter creates an absorption feature around 1400nm, and the overall curve rises into the near-infrared like real soil spectra do. The nutrient values are correlated with soil type and organic matter content (e.g. nitrogen tracks organic matter via the Bremner relation, podzols are acidic, chernozems are nutrient-rich).

What's in it

4,000 samples across 8 soil types
24-band reflectance spectra from 350–1500nm (50nm steps)
Soil chemistry: N, P, K, organic matter, pH
Metadata: soil type, region, sampling depth, texture, land use, crop

Soil types

Soil type	Count	Typical region	pH range	Notes
Chernozem	800	Weinviertel	6.2–7.5	Dark, fertile, high OM
Brown earth	600	Waldviertel	5.5–6.8	Mixed forest/ag soils
Rendzina	500	Alpine foothills	7.0–8.2	Shallow, calcareous
Alluvial	500	Danube valley	6.5–7.8	River-deposited, variable texture
Podzol	400	Alpine	3.8–5.2	Acidic, high OM, low P/K
Gleysol	300	Neusiedlersee	5.0–7.0	Waterlogged, clay-rich
Sandy	400	Marchfeld	5.5–7.0	Light, low OM, droughty
Loess	500	Lower Austria	6.8–8.0	Wind-deposited silt, vineyards

Spectral physics

Each soil type has a characterisitc base reflectance curve shaped by:

Albedo: sandy soils are bright (~~0.18), podzols are dark (~~0.06)
Slope: reflectance generally increases with wavelength into NIR
Organic matter absorption: dip near 1400nm from O-H and C-H bonds, stronger in high-OM soils
Iron oxide absorption: subtle dip around 900nm
NIR plateau: slight bump around 950nm typical of mineral soils

Individual samples vary through Gaussian noise, smooth sinusoidal perturbations, and parameter shifts tied to depth and organic matter content.

Usage

from datasets import load_dataset

ds = load_dataset("Laborator/soil-spectral-analysis")

# Get a sample
sample = ds["train"][0]
print(sample["soil_type"], sample["nitrogen_pct"])

# Quick scatter
import matplotlib.pyplot as plt
import numpy as np

fig, ax = plt.subplots()
for sample in ds["train"].select(range(100)):
    ax.plot(sample["wavelengths_nm"], sample["reflectance"], alpha=0.3)
ax.set_xlabel("Wavelength (nm)")
ax.set_ylabel("Reflectance")
plt.show()

Files

data/soil_spectral.jsonl — one JSON object per line
data/soil_spectral.parquet — same data in Parquet format

Intended use

Training regression models to predict soil nutrients from spectral data
Benchmarking spectral feature extraction methods
Teaching/demos for precision agriculture concepts

This is synthetic data — don't use it for actual agronomic decisions. Real soil spectra have more complexity (moisture variation, surface roughness, mixed pixels). But the spectral shapes and correlations should be good enough for model development and prototyping.

License

Apache 2.0

Downloads last month: 10