.png){kind=icon}
.png){kind=icon}
Return
CIDA: A context-informed decoupling approach for soil trace element estimation using spaceborne cross-sensor data
.png){kind=icon}
PRE
AI
DOI：10.1016/j.isprsjprs.2026.06.007.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Data-driven soil mapping is often limited by small-sample constraints, arising from an imbalance between sparse in-situ measurements and high-dimensional spectral feature spaces. This imbalance increases the risk of overfitting in complex models, restricting their ability to isolate trace metal signals from dominant background soil variability. To tackle these limitations, we developed Context-Informed Decoupling Analysis (CIDA), a mechanism-guided framework intended to separate learnable Zn- and Cu-related spectral structure from dominant background variability and to test its stability across sensors. Using Zhuhai-1 hyperspectral data alongside matched Landsat-8 and Sentinel-2 imagery, we integrated context variables with generalized additive models and parsimonious feature selection to separate Zn and Cu responses from background variability associated with organic carbon and pH. We further evaluated model generalization through random and spatial validation, cross-sensor comparison, network analysis, and MRI-based held-out hotspot prioritization. CIDA improved predictive accuracy under the primary random outer-fold validation setting and preserved a larger share of predictive signal under spatial blocking. In Zhuhai-1, the best CIDA model achieved OOF R2 values of 0.631 for Cu and 0.651 for Zn under random validation, while group-mean retention (defined as the ratio of spatial to random R2 ) under spatial blocking increased from 9.54% to 33.93% for Cu and from 12.67% to 41.68% for Zn relative to conventional strategies. Network analysis revealed a more focused, interaction-based structure for Zn, whereas Cu was associated with a more dispersed, multi-variable topology. Across sensors, Zhuhai-1 generally preserved the strongest predictive structure, further suggesting that continuous spectral shape information can be more critical than broader wavelength coverage when the underlying mechanism is coherent. Furthermore, under spatially blocked held-out assessment, the MRI retained meaningful ranking and hotspot identification capabilities, particularly for Zn. CIDA thus provides a transparent and mechanism-guided framework for constructing robust geospatial models in regions with sparse in-situ data.
Journal
IF:
12.2
Papers:
4.4K
Citations:
3.2W
View more organizations