Extracting Topsoil Information from EM38DD Sensor Data using a Neural Network Approach

Electromagnetic induction soil sensors are an increasingly important source of secondary information to predict soil texture. In a 10.5‐ha polder field, an EM38DD survey was performed with a resolution of 2 by 2 m and 78 soil samples were analyzed for sub‐ and topsoil texture. Due to the presence of former water channels in the subsoil, the coefficient of variation of the subsoil clay content (45%) was much larger compared with the topsoil (13%). The horizontal (EC a –H) and vertical (EC a –V) electrical conductivity measurements displayed a similar pattern, indicating a dominant influence of the subsoil features on both signals. To extract topsoil textural information from the depth‐weighted EM38DD signals we turned to artificial neural networks (ANNs). We evaluated the effect of different input layers on the ability to predict the topsoil clay content. To identify the response of the topsoil, both EM38DD orientations were used. To examine the influence of the local neighborhood, contextual EC a information by means of a window around each soil sample was added to the input. The best ANN model used both EC a –H and EC a –V data but no contextual information: a mean squared estimation error of 2.83% 2 was achieved, explaining 65.5% of the topsoil clay variability with a variance of 0.052% 2 So, with the help of ANNs, the prediction of the topsoil clay content was optimized through an integrated use of the two EM38DD signals.