Improving in-situ spectral estimation of wetland soil organic carbon by integrating multiple optimization strategies

Carbon in wetlands is a critical component of the terrestrial carbon pool in the global carbon cycle. Accurate quantification of wetland soil organic carbon (SOC) is essential for sustainable wetland management and climate change mitigation. However, this effort is often hampered by the inaccessibility of wetlands and high economic and labor costs associated with data acquisition. The advent of visible and near–infrared (Vis-NIR) spectroscopy offers a time- and cost-effective method for monitoring SOC content. Considerable efforts have been directed toward estimating wetland SOC using Vis-NIR spectroscopy. However, great differences were reported for the performance of different predictive strategies, which pose confusion in the choice of methods when using Vis-NIR spectroscopy to estimate SOC. Therefore, this study aims to improve the in-situ spectral estimation accuracy of wetland SOC by testing and evaluating 240 optimization strategies which integrate different pre-treatment technologies, moisture removal algorithms, spectral feature selection methods. Our findings indicate that the optimal strategy involved the combination of first-order derivative (1stD), external parameter orthogonalization (EPO), modified greedy feature selection (MFGS), and one-dimensional convolutional neural network (1D-CNN), yielding R², RPD, RMSE values of 0.94, 4.15, and 1.09 g/kg, respectively. For pre-treatment technologies, 1stD outperformed Savitzky-Golay (SG) and log(1/reflectance) (Log(1/R)). EPO emerged as the optimal moisture removal algorithm, effectively reducing the interference of moisture and other environmental factors. Among spectral feature selection methods, the competitive adaptive reweighting algorithm (CARS) outperformed Boruta, MFGS, and interval random frog (IRF). In terms of prediction models, the 1D-CNN model significantly outperformed partial least squares regression (PLSR) and random forest (RF), with markedly higher accuracy (averaged R² = 0.83, RPD = 2.74, RMSE = 1.76 g/kg). Findings from this research could guide future studies in selecting the optimization strategies to more accurately estimate wetland SOC via Vis-NIR spectroscopy.