This work presents a UAV-based hyperspectral imaging pipeline for high-resolution Soil Organic Carbon (SOC) mapping at the field scale. The system combines hyperspectral remote sensing, ground-truth soil sampling, and machine learning to generate detailed SOC distribution maps that can support precision agriculture and soil monitoring applications.
The pipeline begins with data acquisition using a hyperspectral camera mounted on an unmanned aerial vehicle (UAV). In this study, a real-time hyperspectral imager based on a large Fabry-Pérot Interferometer (FPI) filter was used. Unlike conventional fixed-band hyperspectral systems, the tunable FPI design developed by VTT allows the user to programmatically select the spectral bands and wavelength combinations most suitable for a specific application, enabling more flexible and efficient data acquisition. The sensor operates in the 650–1200 nm wavelength range using an InGaAs detector with 1280 × 1024 pixel resolution and captures full 2D spatial information for every spectral channel while minimizing acquisition delays.
The UAV captures hyperspectral imagery at sub-meter spatial resolution, allowing detailed observation of soil variability across agricultural fields. Compared to conventional multispectral imagery, hyperspectral data provides significantly richer spectral information, enabling the detection of subtle soil characteristics and absorption features related to organic carbon content.
To create reference data for model development, soil samples were collected from multiple locations within each field and analyzed in the laboratory to determine SOC content. The laboratory measurements were then spatially linked with the hyperspectral imagery to form a training and testing dataset consisting of spectral signatures and corresponding SOC values. Image preprocessing included radiometric correction, geometric alignment, and bare-soil identification to reduce the influence of vegetation and non-soil materials on the spectral signal.
Three machine learning algorithms were evaluated for SOC prediction, including Convolutional Neural Networks (CNN), Random Forest, and Gradient Boosting (XGBoost) models. Model performance was assessed using 5-fold cross-validation with metrics such as Root Mean Squared Error (RMSE) and coefficient of determination (R²). Among the evaluated approaches, XGBoost achieved the best predictive performance and was selected for final SOC map generation.
The trained model was subsequently applied to the full hyperspectral dataset to produce continuous high-resolution SOC maps for multiple agricultural fields. The resulting maps reveal fine-scale spatial variability in soil carbon content that would be difficult to capture using traditional sampling approaches alone.