Multi-Response Modeling for Soil Visible-Near Infrared Reflectance Pattern Prediction.

Soils are underappreciated natural resources but its condition are inseparable from the major global issues including the food security and climate change. The status of soil properties, such as carbon and macronutrients, are important to be included in the land resources decision-making processes. However, due to the high cost of soil monitoring, such information is often neglected or at best assumed to be constant. Furthermore, the advances in soil spectroscopy have proven to become an effective solution to reduce the cost of traditional soil laboratory analysis, while maintaining the measurement accuracy. However, the remaining question is how to extrapolate point based soil spectroscopy measurement to the landscape scale that is applicable for the decision-making processes. In this research, we introduce a new pathway of soil prediction using a multiple response modeling technique to predict the soil Visible-Near Infrared (VNIR) spectrum. This soil spectrum is a set of the proportion of the reflected light, i.e. reflectance, that is measured at at 350-2500 nm wavelength region. Therefore, this spectrum represents multiple variables and contains information of multiple soil properties. We utilized the Redundancy Analysis to develop a factorial model for soil VNIR spectrum predictions following the STEP-AWBH soil-factorial modeling framework. Furthermore, we used the soil spectral database from the Florida Soil Carbon Project (N=898, data year 2008/2009) as the model response, and acquired the readily-available global soil forming factor dataset from multiple online sources as the explanatory variables. The Redundancy Analysis yielded predicted spectra with the Root Mean Squared Error (RMSE) 13.8%. The soil properties prediction RMSE using the predicted spectra was 8.5% and was not significantly different compared to the prediction using the measured spectra (p=0.64). This result, therefore, shows that this new technique could be potentially useful to extrapolate the point-based soil spectroscopy measurement to support quantitative land resources management.