Mapping Soil Salinity Combined Sentinel-2 Time-Series Data Mining and Subregional Modeling

Abstract

High-precision digital soil salinization mapping is crucial for agriculture and ecological security. Observation using time-series data improves the accuracy of soil salinization estimation, but how to extract effective information from time-series data is the key to improve model efficiency. This study introduces a high-precision digital soil mapping method by integrating time-series data mining with subregional modeling. First, the study area was divided into vegetated and bare soil zones using NDVI thresholds. The correlation between Sentinel-2 time-series data and soil salinity for both zones was analyzed over the period of 2016.7-2021.8. Second, the optimal time windows for both zones were identified based on the interannual fluctuation range of the optimal temporal phase. Then, feature selection methods were employed to identify key remote sensing variables within the optimal time windows. Finally, the CNN-LSTM model was constructed to predict soil salinity based on the optimal temporal features for vegetated and bare soil zones, leading to high-precision digital soil mapping. The results reveal a distinct annual periodicity in the correlation between remote sensing factors and soil salinity, with an optimal temporal phase each year. Moreover, the correlation shows a decreasing trend with the prolongation of time gap from soil sampling. The optimal time window for the two zones is very different: July to September in vegetated zone and November to January in bare soil zone. This research also found the subregional modeling approach significantly improved prediction accuracy compared to the global model, with R² increasing from 0.72 to 0.89, a 23.61% improvement. Moreover, the temporal mining strategy proposed in this study reduced the data volume by over 98% while enhancing prediction accuracy (R² improved from 0.59 to 0.89, a 50.85% increase). The method, based on time-series feature mining and subregional remote sensing modeling, offers a novel approach for high-precision digital soil mapping.

Keywords: Soil salinization, Time-series data, Subregional modeling, Time-series feature mining, digital soil mapping