Hyperspectral Drone Imagery: Bare Soil Fields

This dataset contains georeferenced tiff images derived from drone-based hyperspectral imaging over bare soil agricultural fields.

The fields Caritas, L1, L3, M5B, R6, RvK1A_B ans S24 were collected during September and October 2025, using an unmanned aerial vehicle (UAV) equipped with a hyperspectral sensor capturing six specific near-infrared bands: 670nm, 730nm, 800nm, 930nm, 970nm, and 1100nm.

The raw sensor data underwent processing using OpenDroneMap software for image stitching, followed by georeferencing correction to ensure spatial accuracy. The resulting tiff files contain digital numbers (DN values) for each spectral band.This dataset may be valuable for agricultural research, soil property analysis, remote sensing development, and related environmental monitoring applications.

Pixel-wise predictions of soil organic carbon (OC, % w/w) for eight agricultural fields in the Ghent region, Belgium (EPSG:32631), derived from UAV-acquired hyperspectral imagery corrected to surface reflectance using calibration panels. Three machine learning models were trained on ground-truth OC measurements from 90+ soil sample points using six spectral bands, their first-order derivatives, and normalised band differences as input features: a 1D Convolutional Neural Network (CNN), a Random Forest (RF), and an XGBoost (XGB) regressor. Models were evaluated via 5-fold cross-validation. One predicted OC map (GeoTIFF) per field per model is provided (fields: B2, Caritas, L1, L3, M5B, R6, RvK1AB, S24).

Two additional training strategies were explored to assess whether incorporating spatial neighbourhood context around each soil sample point improves OC prediction accuracy.

Approach 1 – Aggregated features: For each sample point, pixel reflectance values within a circular neighbourhood of radius r were aggregated into per-band means and standard deviations, yielding one feature vector per sample point. Spectral indices (first-order derivatives and normalised band differences) were then computed on the mean bands. Random Forest and XGBoost models were trained using standard 5-fold cross-validation. Radii of 0.25 m, 0.5 m, and 1.0 m were evaluated during training.

Approach 2 – Pixel augmentation: All pixels falling within a circular neighbourhood of radius r around each sample point were extracted individually, each inheriting the OC label of that plot. This increases the effective training set size proportionally to the neighbourhood area. To prevent data leakage, GroupKFold cross-validation was applied with sample-point identity as the grouping variable, ensuring all pixels from the same plot remain in the same fold. Random Forest, XGBoost, and CNN models were trained under this scheme. Radii of 0.25 m, 0.5 m, and 1.0 m were evaluated during training.