Effects of freezing and soil water content on dissolved organic carbon retrieval precision in the black soil region of Northeast China

Rapid monitoring of dissolved organic carbon (DOC) content is often required in the frozen black soil. It is crucial to assess the soil fertility under different conditions. Nutrient migration and transformation can also be used to elucidate the carbon cycling in the field. Hyperspectral remote sensing has been widely applied to monitor the soil organic matter, due to its rapidity and efficiency. However, the soil freezing can alter the spectral characteristics. It also remains unclear on the accuracy of inversion models under frozen conditions. In this study, a hyperspectral inversion model was established for the DOC content in the frozen black soil. A systematic comparison was made on the accuracy differences of the models over various moisture gradients under frozen versus unfrozen states. Typical seasonally frozen black soil was selected from the Heshan Farm in Heilongjiang Province, China. Random samples were collected with three DOC levels. The DOC content was then measured. Pure water was added at different gradients after measurement. Ultimately, 120 experimental samples were obtained with the moisture contents ranging from 3.57% to 30.14% and DOC concentrations between 171.80 and 322.75 mg/kg. Surface hyperspectral reflectance was collected under unfrozen and frozen states using an AvaField-3 field spectroradiometer. Five spectral preprocessing approaches were applied: raw spectral reflectance (REF), first-order differential reflectance (FDR), second-order differential reflectance (SDR), logarithm of reciprocal (LR), and standard normal variable (SNV). Variable importance in projection (VIP) was used to select the sensitive bands. A one-dimensional convolutional neural network (1D-CNN) was employed for the feature extraction.The dataset was partitioned using the Kennard-Stone (KS) algorithm with a calibration-to-validation ratio of 2:1, resulting in 80 samples for the calibration set and 40 for the validation set. The DOC inversion models were constructed using four machine learning models, including backpropagation neural network (BPNN), random forest (RF), tabular prior-fitted network (TabPFN), and extreme gradient boosting (XGBoost). The optimal model was selected from the low and high moisture gradient categories. The performance of the model was finally evaluated using the coefficient of determination for calibration (R_C²), coefficient of determination for prediction (R_P²), root mean square error (RMSE), and ratio of performance to deviation (RPD). The results indicated that: 1) All four models effectively predicted the soil DOC content. The accuracy was ranked in descending order under both unfrozen and frozen states: RF > XGBoost > TabPFN > BPNN. The RF-FDR model demonstrated the best performance was achieved (unfrozen: R_P²=0.851, RMSE=13.095 mg/kg, RPD=2.410; frozen: R_P²=0.808, RMSE=14.830 mg/kg, RPD=2.123). While the BPNN-LR model demonstrated the worst performance (unfrozen: R_P²=0.742, RMSE=17.200 mg/kg, RPD=1.889; frozen: R_P²=0.687, RMSE=18.945 mg/kg, RPD=1.670). 2) Soil freezing status significantly dominated the accuracy of DOC inversion among models. Comparative analysis revealed that the optimal preprocessing remained consistent before and after freezing. However, the freezing reduced the inversion accuracy of all models, with the most substantial decline in BPNN (R_P² and RPD decreased by 7.36% and 11.62%, respectively), and the least in XGBoost (R_P² and RPD decreased by 4.70% and 7.25%, respectively). 3) Moisture content gradients considerably influenced the accuracy of the model. The high-moisture models outperformed the low-moisture ones under unfrozen conditions, with R_P² and RPD higher by 0.067~0.082 and 0.206~0.392, respectively. There was a decrease in the accuracy of both low-moisture and high-moisture models under frozen conditions. The low-moisture models were reduced significantly (R_P² and RPD decreased by 16.67%-17.71% and 13.76%-18.69%, respectively). The RF-FDR model can also provide the technical support for the hyperspectral monitoring of the DOC content in seasonally frozen black soil.