Enhancing proximal and remote sensing of soil organic carbon: A local modelling approach guided by spectral and spatial similarities

• The local modelling approach improved SOC prediction for three different spectral datasets. • The proposed method searches spectrally and spatially similar samples for localized SOC prediction. • SOC prediction was improved because soil compositional heterogeneity was reduced in the training set. • SOC prediction models based on soil RGB imagery had the biggest level of improvement. • Cross-scale transferability of the local modelling approach was independently validated. Optical proximal and remote sensing of soils capture essential spectral features needed to develop spectra-based prediction models for soil organic carbon (SOC), which determines a variety of biogeochemical processes on Earth. At large scale, increasing soil heterogeneity complicates the response relationship between soil spectra and SOC, making global models ineffective for local SOC predictions. Here, we propose a local learning approach that searches spectrally and spatially similar samples for site-specific SOC predictive modelling. The approach was developed using three large-scale spectral datasets (n = 911), including soil RGB images acquired by a smartphone, laboratory visible and near-infrared spectra using diffuse reflectance spectroscopy (DRS), and remotely sensed bare soil spectra (RSBS) from Sentinel-2 imagery, and further validated against three independent datasets of the same types (n = 72) at catchment scale. Different spectral similarity metrics, and weighted combinations of spectral and geographical similarity matrices were tested to optimize the selection of local training samples. As a result, the optimal modelling strategy, with partial least squares regression (PLSR) as the local fitting algorithm, consistently produced superior performances (R 2 : 0.66 to 0.82) than the conventional global modelling approach (R 2 : 0.59 to 0.77) for all three data types. Increasing the weight assigned to geographical similarity during local learning enabled the selection of compositionally similar samples, which improved SOC prediction accuracy by 10.87 %, 2.50 %, and 7.21 % for RGB, DRS, and RSBS datasets compared to models relying on spectral similarity alone. Independent validation shows that the performance of global PLSR models deteriorated from regional to catchment scale, while the proposed local modelling approach could maintain the prediction accuracy for all three data types (R 2 : 0.64–0.80), proving its cross-scale transferability. These findings highlight the potential of integrating spectral and spatial information to optimize the local learning procedure for accurate SOC quantification and mapping, especially when dealing with complex datasets covering large spatial scales.