Systematic comparison of band selection strategies and sensor configurations for soil moisture classification using UGV-mounted hyperspectral imagery

Accurate soil moisture monitoring is essential for precision agriculture, but current remote sensing and in situ sensing approaches are often constrained by fixed installations, limited sensing depths, or delayed measurements. Hyperspectral sensors deployed on unmanned ground vehicles (UGVs) provide flexible mobile platforms for real-time, high-resolution sensing. This study presents a systematic comparison of band-selection strategies and sensor configurations for soil moisture classification under realistic field conditions. Soil samples were collected throughout the growing season from three agricultural fields in North Dakota, USA. Hyperspectral imagery was acquired using two sensors, SPECIM FX10 (400–1000 nm) and FX17 (900–1700 nm), and soil moisture was grouped into four classes based on gravimetric measurements collected at depths of 6, 12, 18, and 24 inches. Informative spectral bands were identified using Random Forest (RF) and Partial Least Squares Regression (PLSR). Feature-guided band selection improved classification performance compared with evenly spaced or consecutive band subsets, with RF-based selection yielding the strongest performance with reduced bands. The most informative wavelengths were concentrated in the shortwave infrared (SWIR) region, particularly around 1400–1500 nm, corresponding to strong water absorption features. Sensor configuration analysis showed that the FX17 sensor achieved higher and more consistent accuracy (0.9583) across all soil depths (6–24 inches) compared with FX10 alone (0.8889) and the class-matched cross-sensor spectral combination (0.9167). Because the spectra were matched by moisture class rather than acquired from the same soil sample, the combination should be interpreted as a comparison of spectral coverage rather than as strict sensor fusion. Among the evaluated deep learning models, the one-dimensional convolutional neural network (CNN) achieved the best overall performance compared with the long short-term memory (LSTM) and CNN–LSTM architectures, while also maintaining a simpler structure and lower computational complexity. The optimal CNN model trained on the FX17 dataset was further applied to generate image-level soil moisture classification maps, demonstrating the scalability of the proposed framework for spatial soil moisture mapping. Overall, the results indicate that feature-guided band-selection, appropriate sensor configuration, and SWIR-rich hyperspectral sensing can effectively support soil moisture classification for field monitoring and precision agriculture.