A Propagation Model for Millimeter-Wave Radar Performance Assessment in Agricultural Scenarios

Abstract

Grain carts decouple a combine's harvesting rate from a truck's hauling schedule, but cart operators currently have only a rough, largely visual sense of how full the cart is at any given moment. A millimeter-wave (mmWave) radar mounted above the cart and scanning down across the bin as it fills could provide a continuously updated, through-dust estimate of pile height and shape, but the radar return in this environment is shaped not only by the grain pile itself but by the cart's canopy bars, the falling grain stream from the unloading auger, and airborne dust, all of which must be understood before a fill level estimate can be trusted. This thesis develops a physics-based propagation and signal simulation of a grain cart filling scene and validates the simulated radar returns against measured data from a real grain cart. The simulation models an FMCW phased array radar illuminating a time-evolving scene composed of a Gaussian height field grain pile, three fixed cylindrical canopy bars, a falling grain chute, and a volumetric dust cloud, with line-of-sight self-shadowing of the pile computed at every time step. An antenna self-localization algorithm was developed that exploits the bars' known geometry as fixed reference reflectors, allowing the system to estimate its own antenna position from the radar returns themselves rather than assuming a perfectly known, static mount. Comparing simulated and measured range and azimuth imagery shows that the model reproduces the dominant geometric behavior of the real scene. The canopy bar returns appear at matching locations in both datasets, and the pile return broadens and shifts as fill level increases in a manner consistent with the simulated line-of-sight obstruction model, which shows a two-phase pattern: visibility of the pile first rises as a small pile comes into clearer view of the antenna, then falls as a larger pile increasingly shadows itself. The comparison also identifies the model's largest gap: an elevated, structured noise floor in the measured data, attributable to un-modeled backscatter from the cart's metal floor and walls, is shown to be a more significant source of simulated-versus-measured mismatch than dust, the falling grain return, or the canopy bars, each of which is also shown to differ from the real scene in specific, explainable ways. The antenna self-localization algorithm achieves sub-meter three-dimensional position accuracy on average, with no evidence of cumulative drift, though accuracy is uneven across axes. Together, these results demonstrate that a bar-anchored, physics-based propagation model can capture the key geometric behavior of a dynamic, cluttered agricultural radar scene, while also identifying adding cart floor and wall backscatter to the signal model as the clearest priority for closing the remaining gap between simulation and measured reality.