The validity domain of sensor fusion in sensing soil quality indicators

Soil health has gained increasing attention under the rapid development of industrialization and the requirement for green agriculture. Therefore, up-to-date soil information related to soil health is urgently needed to ensure food security and biodiversity protection. Previous studies have shown the potential of proximal soil sensing in measuring soil information, while it remains challenging to get cost-efficient and robust estimates of multiple soil health indicators simultaneously via sensor fusion. In this study, we investigated the potential of visible near-infrared (vis-NIR), and mid-infrared (MIR) spectroscopy as well as three model averaging methods in predicting three soil health properties, including soil organic matter (SOM), pH, cation exchange capacity (CEC). The model averaging methods are not only used for model fusion but also for high-level sensor fusion, which include Granger-Ramanathan (GR), Bayesian Model Averaging and Spectral-Guided Ensemble Modelling (S-GEM). Here, S-GEM is a recently proposed algorithm that can improve soil spectroscopic prediction by including spectral information in ensemble modelling. Four widely used prediction models were evaluated, including partial least square regression, Cubist, memory based learning and convolutional neural network. For SOM, sensor fusion based on model averaging algorithms was comparable to that of Sensor(single) + Model(multiple) (MIR singly based on S-GEM) with R-2 of 0.86. However, MIR only with S-GEM performed the best among all methods (LCCC of 0.92, RMSE of 3.66 g kg(-1) and RPIQ of 3.68). The 10-fold cross-validation results indicated that Sensor(single) + Model(multiple) (MIR singly based on S-GEM) performed best among all methods for pH, with R-2 of 0.84, LCCC of 0.90, RMSE of 0.45 and RPIQ of 3.65. For CEC, Sensor(multiple) + Model(multiple) based on GR performed best with R-2 of 0.66, LCCC of 0.80, RMSE of 3.48 cmol + kg(-1) and RPIQ was 2.22. Our results also showed that sensor fusion failed to improve spectral prediction of soil information when the performance among sensors differed a lot (Delta R-2 > 0.2), and the use of a single best sensor is therefore suggested in this case. When the sensors have a close model performance (Delta R-2 < 0.2), Sensor(multiple) + Model(multiple) based on GR was recommended. The outcome of this study can provide a reference for determining the validity domain of sensor fusion methods in improving the accuracy of soil health prediction.