An auto-tuning process-based 3D model to forecast irrigation demand in kiwifruit

In a contest of climate change and increasing world population, the optimization of agricultural inputs such as water and fertilizer is of utmost importance to assure crop quality and to limit the impact of agriculture. For these reasons, the need for robust methods of agricultural modeling and forecasting has never been clearer. In particular, forecasting the water budget is a key tool for reducing water wastage and maximizing agricultural production. Although process-based models are largely used in off-line simulation and studies, their operational use in forecasting the irrigation requirements of a specific crop remains complex and the level of accuracy achieved is often insufficient since, if used alone, process-based simulation systems fail to capture the soil and plant dynamic behaviors. To overcome these limitations we propose an integrated system coupling Orchard3D-Lab, an innovative three-dimensional process-based model specifically devised for fruit trees, with a state initialization procedure that exploits a two-dimensional probe grid. The system is capable of auto-tuning its parameters on a specific soil and of providing a precise forecast that can support precision watering policies on a weekly horizon. A large set of tests has been conducted on Kiwifruit in an experimental farm in Northern Italy. Besides accuracy, tests proved the robustness of the system even in the presence of a limited set of examples for parameter auto-tuning. This makes our approach concretely applicable in real-world settings. • Orchard3D-lab, an innovative 3D process-based model to compute the soil water balance of an orchard. • A State Initialization technique to map coarse-grained 2D/3D-sensor grid to fine-grained profiles. • A Parameter Tuning technique to automatically calibrate the model parameters. • A detailed field experiment in a commercial Kiwifruit orchard.