Soil salinity inversion by combining multi-temporal Sentinel-2 images near the sampling period in coastal salinized farmland

Chongxin Duan ¹ Yong Zhang ² Chaopu Hu ² Hongyan Chen ^1* Peng Liu ³

• ¹ National Engineering Research Center for Efficient Utilization of Soil and Fertilizer Resources, College of Resources and Environment, Shandong Agricultural University, Tai'an, China
• ² Shandong Territorial Spatial Planning Institute, Jinan, Shandong Province, China
• ³ College of Agriculture, Shandong Agricultural University, Tai'an, China

Rapid and accurate soil salinity (SS) analysis is essential for effective management of salinized agricultural lands. However, the potential of utilizing periodic remote sensing satellite data to improve the accuracy of regional SS inversion requires further exploration. This study proposes a novel inversion approach that combines multi-temporal images captured near the SS field sampling period (September 5-10, 2020). Focusing on Wudi County, China, we analyzed three time-series Sentinel-2 images obtained near the sampling period to determine the inversion time window. Images within the window were synthesized into four combined-temporal images through three arithmetic operation strategies and one band combination strategy. SS-related spectral variables derived from both single and combined-temporal images were selected using Random Forest (RF), ReliefF, and Support Vector Machine Recursive Feature Elimination algorithms (SVM-RFE). Subsequently, inversion models were developed and compared using an Extreme Learning Machine. The optimal model was then applied to map regional SS distribution. The results demonstrate that: (1) combinedtemporal models consistently outperformed single-temporal models, particularly those employing the band combination strategy, showing a 0.25-0.53 higher mean Relative Percent Deviation (RPD); (2) models utilizing RF for variable selection exhibited superior stability and efficiency, with a mean RPD 0.02 to 0.04 higher than models using other algorithms; (3) the ELM model with band combination image and RF variable selection achieved the highest validation precision (Coefficient of Determination = 0.72, Root Mean Square Error = 0.87 dS/m, RPD = 1.93); (4) the final SS inversion map revealed a spatial gradient of increasing salinity in farmland from the southwestern area toward the northeastern coastal region, with 46.7% of farmland exhibiting yield-affecting salinity levels. These findings provide empirical insights into the development of soil remote sensing techniques and supporting agricultural-environmental management strategies.