Physics-guided transfer from laboratory spectra to satellite observations reveals coupled salinity–texture–fertility variability in arid irrigated croplands

Satellite prediction of multiple soil attributes in arid irrigated croplands is limited by the spectral-domain gap between proximal spectra and satellite observations. Here, we developed a hierarchical, physics-guided, multimodal, multitask network (HPM 3 -Net) for laboratory–field–satellite soil spectral transfer. Using 1,250 topsoil samples from the Manas River Basin, laboratory spectra, field spectra, ZY1 Advanced Hyperspectral Imager (AHSI) data and Sentinel-2 (S2) observations, we predicted 12 soil attributes related to salinity–alkalinity, fertility and texture. The framework integrated cross-modal alignment, teacher–student distillation, hierarchical multitask learning, sensor-response-function (SRF) harmonization, and constraint-based regularization to improve the plausibility and internal consistency of multi-property predictions. Laboratory and field spectra were harmonized to satellite bands using SRF convolution rather than treated as direct substitutes for surface reflectance. The HPM 3 -Net achieved an overall mean R 2 of 0.755 ± 0.011. The 2025 temporal validation reduced the overall mean R 2 by 10.55%. The sensitivity analysis showed that the overall mean R 2 remained between 0.730 and 0.755. More importantly, satellite-scale observability, defined here as the extent to which a soil attribute retains stable spectral–spatial signals that can be sensed after transfer from laboratory spectra to satellite observations, was process-dependent. Texture fractions, soil organic matter, total salt content and calcium carbonate equivalent were more strongly expressed through stable spectral–spatial signals, whereas fertility-related attributes and some alkalinity indicators depended more on indirect coupling with soil texture, salinity and management-related soil conditions. These findings highlight the benefits and limits of physically constrained spectral transfer for satellite-domain soil prediction.