Abstract: Soil water content, as one of the most critical indicators of coastal wetland ecosystem health, consistently plays the leading role in the growth and invasion dynamics of Spartina alterniflora. Moreover, S. alterniflora invasion reciprocally induces substantial modifications in soil conditions, particularly soil water content, thereby exerting profound ecological impacts on wetland systems. Consequently, water content patterns and their predictions for soils at different depth levels in such environments may exhibit distinctive characteristics because of the interdependent relationships. This study developed an effective prediction framework integrating Boruta algorithm and random forest regression with visible (350-700 nm), near infrared (700-1 100 nm) and shortwave infrared (1 100-2 500 nm) spectroscopy to systematically monitor soil water content during S. alterniflora invasion in coastal wetlands. The investigation was conducted in a S. alterniflora wetland within China’s Yancheng National Nature Reserve, Jiangsu Province, where ecological invasion has persisted for decades. The sampling sites was selected according to the stratified random sampling method, and the sample soils was collected from 5 different depth levels (0-10 cm, >10-20 cm, >20-30 cm, >30-60 cm and >60-100 cm). A total of 240 soil samples were collected from 48 sites, well preserved and sent to laboratory. The reflectance spectra and water content of these samples were then measured and acquired for all samples, enabling the analysis of variations in both water content and reflectance between samples at different depth levels. The Boruta algorithm effectively identified spectrally sensitive wavelengths, which were subsequently implemented as predictive variables in random forest regression models. The performance of the comprehensive prediction model utilizing the complete dataset and stratified prediction models calibrated with samples at single depth level was evaluated and compared afterwards. Analytical results revealed a characteristic distribution pattern that the soil water exhibited progressive depletion with increasing depth, except for the surface depth (0-10 cm), while spectral reflectance demonstrated consistent elevation across deeper soil samples. This inverse relationship can be explained by the negative correlation between reflectance and water content for the absorption of water across Vis-NIR-SWIR range, which was confirmed in this research as well through Pearson’s correlation analysis, revealing maximal correlation (r=−0.89) at 1 910 nm wavelength. The stratified random forest prediction models calibrated with samples at single depth level demonstrated declining predictive accuracy with increasing soil depth, achieving optimal performance at 0-10 cm (R²=0.95) and poorest performance at >60-100 cm (R²=0.87). The comprehensive prediction model incorporating all samples yielded intermediate accuracy (R²=0.913, RMSE=0.053), outperforming stratified models for >30-60 cm and >60-100 cm depths while equaling the prediction model for >20-30 cm depth. This enhanced performance might derive from the expanded training dataset utilized by the comprehensive prediction model for all depth levels, while the stratified prediction models utilized dataset at specific depth level. Though, the stratified prediction models still showed the advantage at depth level of 0-10 cm, >10-20 cm and >20-30 cm, reaching satisfied accuracy with smaller dataset for training. These results demonstrated the potential of random forests modelling in rapid monitoring of soil water content in S. alterniflora wetland, the capacity of Boruta algorithm in selecting sensitive bands from reflectance spectra, and the potential accuracy enhancement through depth-stratified modeling approaches.