Effect of smart phone cameras on color-based prediction of soil organic matter content

Compared with the complicated operation of traditional laboratory methods or expensive spectral instruments, soil organic matter (SOM) content prediction based on smart phone photos has recently received heightened attention. However, as one of the most popular mobile devices, the imaging characteristics of smart phone cameras are quite different due to the differences in manufacturer technologies, which may affect the relationship between the photo colors and SOM content. Whether the highly accurate model built based on a single phone can be applied to other phone types is still an open question. This study has validated the shared capacity of color-based prediction models, analyzed the intrinsic factors affecting the shared capacity, and proposed potential methods to enhance the shared capacity. In total, five smart phones were selected for the study, and dried soil samples were photographed in an optical dark chamber. Imaging spectroscopy was used to scan the samples. The photo and spectral data were pretreated, and stepwise multiple linear regression (SMLR) and partial least squares regression (PLSR) models were built. The spectral response curves of the five smart phone cameras were also obtained separately to clarify their imaging characteristics. Results indicated the RGB color distribution conditions of soil photos obtained by different phones were different, which affects the correlation between the color parameters and SOM. The prediction ability of the models constructed by the five smart phones were similar to the spectral devices, achieving an R-2 of 0.68-0.77 and an RMSE of 5.32-7.12 g/kg. However, when substituting the color parameter datasets obtained by the five phones into the models constructed by the other phones to verify the shared capacity, we found that most of the prediction results could not meet the requirements for use. The poor shared capacity might be extremely disruptive to users. We proposed several potential methods that might enhance the model's shared capacity. The results of this study showed that smart phone cameras have a good capability of modeling SOM content independently, but the shared capacity of different phones still needs further investigation.