Lite-HCAR-CNN: A Lightweight Deep Learning Model for Rapid Forest Soil Organic Carbon Estimation from Laboratory Vis-NIR Spectra

Rapid and reliable monitoring of soil organic carbon (SOC) is important for precision forest management and soil ecological assessment. In this study, 853 surface soil samples were collected from subtropical plantation and secondary forest regions in Guangxi, China. After spectral quality control and SOC outlier screening, 842 valid samples were retained for model development. Laboratory-acquired visible and near-infrared spectra were preprocessed using Savitzky-Golay smoothing and standard normal variate correction, followed by a dimensionality-reduction comparison involving the full spectrum, 10 nm, 20 nm and 40 nm binning, principal component analysis, and PLS-VIP feature selection. To improve computational efficiency for rapid SOC estimation from Vis-NIR spectra, a lightweight model named Lite-HCAR-CNN was developed. The model adopts two one-dimensional convolutional blocks with a kernel size of 7, dual squeeze-and-excitation channel recalibration, Swish activation, and a compact residual dense regression head. Model evaluation employed an SOC-stratified outer five-fold cross-validation strategy, with an inner validation subset used only for checkpoint selection and each held-out outer test fold evaluated once. Under the selected 20 nm binning strategy, Lite-HCAR-CNN achieved an average R² of 0.810 ± 0.027 and RMSE of 4.617 ± 0.373 g/kg. Compared with 1D-CNN, Lite-HCAR-CNN reduced RMSE by approximately 8.4%, and compared with 1D-ResNet18 it reduced trainable parameters by 87.9%, model storage size by 87.9%, FLOPs by 96.7%, and CPU inference latency by 88.6%. Binning comparison, kernel-size and hidden-unit sensitivity analyses, ablation experiments, residual diagnosis, and MC Dropout uncertainty analysis were further conducted to clarify the model design and reliability. The results indicate that Lite-HCAR-CNN achieves a favorable balance between prediction accuracy and computational efficiency for fast estimation of forest soil organic carbon using laboratory Vis-NIR spectroscopy.

Hali tarjima qilinmagan