Soil salinity poses a critical threat to agricultural productivity in arid and semi-arid regions, particularly under changing climate. In Morocco's Sehb El Masjoune area, we hypothesize that post-rainfall terrain dynamics and soil spectral responses drive the spatial variability of surface salinity. This study integrates Sentinel-2, Landsat-9, and PlanetScope imagery with fieldmeasured electrical conductivity and machine learning. It evaluates three hypotheses: (i) micro-topographic depressions retain moisture and promote localized salt accumulation; (ii) soil composition clusters influence differential salt retention; and (iii) combining multi-source data improves salinity mapping. Two novel post-rainfall proxies were developed from PlanetScope imagery. The first, the Depression Proxy (DP), identifies moisture-retaining concavities. The second, the Soil Clusters Proxy (SCP), groups soils based on post-rainfall spectral responses that are linked to texture and moisture properties. These were integrated with spectral indices and terrain variables into three modeling approaches using RF and GBR. Sentinel-2 combined with GBR and DP feature achieved the highest accuracy on an independent test set (R2 = 0.85), identifying concave terrain as persistent salinity location and highlighting the role of surface topography (i.e., micro-depressions) in salinity distribution. Categorical accuracy confirmed that 56 % of samples were assigned to the exact soil salinity class and 80 % within +/- 1 class. Additionally, seasonal changes in predicted salinity were also examined using consistent spectral signatures between wet and dry imagery; however, due to the lack of wet-season ground truth, the resulting map represents qualitative spatial trends rather than validated salinity estimates. This process-informed Earth Observation-based framework improves the accuracy and interpretability of salinity mapping.