Abstract

An increasing number of soil spectral libraries are being developed at larger extents, including at national, continental, and global scales. However, the prediction accuracy of these libraries was often fairly poor when used on local scales. This study evaluates different strategies to improve the model accuracy of a regional spectral library for soil organic carbon prediction in four different local target areas. In total, five strategies using the Partial least squares regression (PLSR) were compared, including the use of local, spiked-regional, spiked-subset-regional and two localized models (memory based learning (MBL) and localized PLSR). MBL derives a new local prediction model based on a subset of the regional spectra similar to the new sample to be predicted. A localized PLSR calibrates a linear regression model using projected scores of the local samples derived from a pre-trained regional PLSR model. Validation results showed that the performances of the spiked models were not much better than those derived from the local and localized models. With >20 local samples, the localized PLSR model performed better than both the local and spiked-regional models. MBL model achieved similar performance to the localized PLSR model. Nevertheless, the accuracy of the models was heavily affected by both the spectral similarity of the data and the statistics of the predictand. Therefore, we recommend localizing the prediction models. Our results also suggest that spiking affected the regression coefficients of the PLSR models but not the loadings, which enabled the compression of spectra data into informative PLS scores. If the local spectra have low similarity to the regional spectral library, building a local spectral library would be more beneficial, assuming the sample size is large enough (>30).

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call