Frieke Vancoillie

A neural network approach to topsoil clay prediction using an EMI-based soil sensor

High-resolution proximal soil sensor data are an important source of information for optimising the prediction of soil properties. On a 10.5 ha arable field, an intensive EM38DD survey with a resolution of 2 m × 2 m resulted in 19,694 measurements of ECa-H and ECa-V. A large textural variation was present in the subsoil due to the presence of former water channels. Nevertheless, both ECa-V and ECa-H data displayed the same spatial variability. This spatial similarity indicated the strong influence of the subsoil heterogeneity on the ECa-H measurements. Using variography, two scales of ECa variability were identified: short-range (∼35 m) variability associated with the channel pattern and wider within-field variability (∼200 m). Using artificial neural networks (ANNs), prediction of the topsoil clay content was optimised (i) by using an input window size of 3, 5, 7, 9, and 11 pixels to account for local contextual influence and (ii) by including both ECa-H and ECa-V in the network input layer to isolate the response from the topsoil. The models were evaluated using R 2 and the relative mean squared estimation error (rMSEE) of the test data. The most accurate predictions were obtained using both orientations of the EM38DD sensor without contextual information (R 2 = 0.66, rMSEE = 0.40). The importance of ECa-V on the topsoil clay prediction was expressed by a relative improvement of the rMSEE of 29%. For comparison, a multivariate linear regression (MVLR) was performed to predict the topsoil clay content based on the two orientations. The ANN models up to a window size of 5 pixels outperformed the MVLR, which resulted in an R 2 of 0.42 and an rMSEE of 0.63. ANN analysis based on both orientations of the EM38DD appears to be a useful tool to extract topsoil information from depth-integrated EM38DD measurements.