Sensor data fusion for topsoil clay mapping

The study investigated proximal sensor data fusion for topsoil clay mapping on a 22 hectare agricultural field in southwest Sweden. Eight different predictor sets and two different prediction methods were tested in an orthogonal design. The predictor sets were different combinations of proximally measured gamma (γ) ray spectrometry and apparent electrical conductivity (ECa), four terrain attributes (elevation, slope and the cosine and the sine of the aspect) and the digital numbers (DNs) of an aerial photo. The two prediction methods were partial least squares regression (PLS-R) and k nearest neighbor prediction (kNN). It was found that the γ ray spectrometry variables (232Th, 40K and total count of decays) were good predictors of topsoil clay content (mean absolute error of about 1.5% clay) and predictions were neither much improved nor deteriorated by addition of any of the other predictors. The ECa measurements, which are affected also by the subsoil, did not perform as well. Predictions were improved when the ECa data were integrated with the aerial photo DN but were deteriorated by addition of elevation data. The kNN method yielded slightly better predictions than the PLS-R method but overall it was more important which input data were used than how the predictions were made. It was observed that even though dense soil sampling was used for calibration (three samples per hectare), use of proximal soil sensor data was almost always better than mere interpolation of the calibration samples.

Figure optionsDownload as PowerPoint slideHighlights
► Sensor data fusion was tested for high resolution mapping of an agricultural field.
► We tested eight predictor data sets and two methods for prediction of topsoil clay.
► Use of ancillary data was often better than spatial interpolation of soil samples.
► Gamma ray spectrometry data were good predictors of topsoil clay content.
► Aerial photo or gamma ray spectrometry data improved predictions based on ECa.