The non-invasive detection of soil compaction and the effect of traffic management on crop performance

The increasing weight of contemporary agricultural vehicles causes risk of widespread soil compaction which is a threat both to productivity and sustainability of soil resources. A range of prevention measures are being introduced in agriculture, of which Controlled Traffic Farming (CTF) and low tyre pressure (LTP) systems are commercially adopted. A quick proximal method to detect the presence and severity of soil compaction would be beneficial as it would enable planning spatially targeted, less expensive and potentially more effective remediation operations. This thesis reports on a 3-year study, where the experimental design was based on precisely planned field traffic consisting of offset tractor passes, so the resulting sub-plot spatial structure consisted of a “stripey” pattern of elongated rectangles that received a known number of wheel passes annually (0-5 passes), in combination with tillage depth and tyre pressure. Supporting studies were conducted on a sandy loam site in West Midlands, UK, with a 10-year history of controlled traffic of similar design, and in the soil hall at Harper Adams University. Soil and crop response to the experimental factors were measured and analysed, testing the application of ground penetrating radar (GPR) for soil compaction detection. Agricultural traffic was found to have an adverse effect on soil physical properties by increasing penetrometer resistance and bulk density, and on crop performance, by decreasing emergence percentage and yield, typically by 20-30% compared to untrafficked areas, with higher losses recorded in winter barley in wet conditions. The application of low tyre pressures was found to increase emergence, NDVI and yield, compared to standard pressures. On the field sites, heavily trafficked tractor wheelways and relatively untrafficked plot centres were scanned with GPR. In the soil hall, three zones were scanned: untrafficked, trafficked, trafficked and covered with loose soil. Scanning was performed statically at 0.4 m height. Numeric signal attributes derived from GPR traces served as feature vectors for supervised classification by the following machine learning algorithms: 1) deep learning (Keras), 2) random forest (RF), 3) the model with the highest AUC supplied by the H2O AutoML function. Training and classification were done within each dataset separately. In the laboratory setting, with 3 predicted classes (traffic zones), overall accuracy was 59% with deep learning and 83% with random forest. In one-vs-one tests, accuracy was 70-84% with deep learning and 83-96% with random forest. In the detection of traffic zone on the field sites, the leader (highest AUC) models selected by the H2O AutoML function, all Stacked Ensemble type, achieved accuracy 67-87%, depending on dataset. A higher-than-random accuracy using Keras and RF was achieved in one dataset only (67% Keras, 71% RF). Accompanying results point to the possible role of traffic-related variability in soil moisture in detection.