Do we really need large spectral libraries for local scale SOC assessment with NIR spectroscopy?

• Different-sized models were used to predict SOC at local scale using NIR spectroscopy.
• These models were adapted to target site characteristics by extra-weighted spiking.
• Once adapted, small-sized models produced accurate SOC predictions (R2 > 0.95; RPIQ > 6).
• The adaptation by extra-weighted spiking allows downscaling and transfer of NIR models.
• Large spectral libraries are not needed to predict SOC at local scale.

Near infrared (NIR) spectroscopy was used to predict the soil organic carbon (SOC) contents at local scale in eleven target sites. For that, eight spectral libraries of different sizes (ranging from 3482 to 36 samples) were used to construct national, provincial and local scale models. Inaccurate predictions were obtained except when the largest national library was used to construct the model. We also obtained SOC predictions once the models were adapted to target sites characteristics. For the models’ adaptation, we used a two-step approach consisting on spiking (as first step) and extra-weighting (as second step). The effect of spiking was small in larger-sized models and high in smaller-sized models, whereas the effect of extra-weighting was small in smaller-sized models and large in larger-sized models. The very high accuracy obtained after models’ adaptation (R2 > 0.95; RPIQ > 5.48), regardless of the size of the spectral library, suggests that large spectral libraries are not needed for local scale SOC assessment. These results have important implications regarding the way that NIR spectroscopy can result highly effective for land management and how users can focus and organize the analytical efforts.