Predicting the spatial distribution of soil salinity based on multi-temporal multispectral images and environmental covariates

• An adaptive region energy weighted pyramid transform was used for image fusion. • Multi-temporal remote sensing images can capture the dynamics of soil salinity. • Adding environmental covariates can slightly improve the accuracy of EC prediction. • Topography and groundwater have proved more useful for EC mapping. Soil salinization hinders global agricultural development and environmental sustainability. Therefore, effective monitoring and assessment of soil salinization are crucial for implementing appropriate mitigation strategies. Optical remote sensing satellites can achieve rapid regional topsoil salinization mapping. However, previous studies rarely considered the soluble salt deposition or crystallization dynamics in soils, especially at large regional scales, where the existence states of soluble salts are uncertain due to environmental differences. As a result, soil salinity mapping based on mono-temporal images exhibits low accuracy and strong contingency. In light of this, multi-temporal image fusion is considered in this study, and an adaptive regional energy weighting Laplacian pyramid transform (ARW-LPT) method is proposed to fuse Landsat-8 images covering the bare soil periods over nearly a decade to fully exploit the spectral characteristics of soil salinity. Considering the different soil salinization processes, mechanisms, and degrees in different regions, various environmental covariates related to soil salinization, such as topography, groundwater, soil, vegetation, and climate, are incorporated to enhance the generalizability and interpretability of the prediction model. The study area is in the western Jilin Province, China, covering 81862.20 km 2 , and 343 soil samples are collected. Feature importance of the fusion bands and environmental covariates is assessed using eXtreme gradient boosting and SHapley Additive exPlanation analysis. A soil electrical conductivity prediction model is established using a convolutional neural network. The results show that 1) the accuracy of soil salinity prediction based on multi-temporal fusion images is higher than that based on mono-temporal images, with an R 2 increase of 0.19, an RMSEP decrease of 3.10 ds·m −1 , and an RPIQ increase of 0.80. 2) Incorporating environmental covariates leads to a better performance in revealing the spatial variability of soil salinity, with the contributions of different environmental covariates to prediction accuracy ranked as topography > groundwater > soil > climate > vegetation > land type. 3) Combining multi-temporal images and environmental covariates leads to the highest soil salinity prediction accuracy and mapping ability, with an R 2 of 0.79, an RMSEP of 5.69 ds·m −1 , and an RPIQ of 2.72. This study highlights the importance of considering the changes in relevant environmental factors and spectral variations during soil salt accumulation, thereby shedding more light on the soil salinization mechanisms and informing soil salinity prediction.