Smart Agricultural Technology (Dec 2023)
Modeling the soil-machine response of secondary tillage: A deep learning approach
Abstract
Seedbed preparation constitutes a highly important step in the crop establishment process since it directly influences both quality and yield expectation in crop fields. The quality of seedbed preparation is usually assessed by the operator of the tractor or a supervising farmer and thus dependent on qualified field personnel which are scarce today. Subsequent to the visual assessment of the current seedbed structure, a qualified operator can manually adapt the secondary tillage machine in order to optimize resource efficiency (e.g., fuel consumption) while maintaining optimal seedbed preparation according to local soil conditions. In this work, we propose to use an automated approach for real-time in-situ measurement of seedbed quality. Stereo cameras mounted in the front and the back of the tractor provide images and depth-information. Additionally, telemetry data such as working speed, power take-off (PTO) speed, fuel consumption and engine torque utilization of the tractor are logged. Based on this data, we develop a Deep-Learning-based model to predict seedbed quality (measured in form of the Roughness Coefficient), engine torque utilization and engine fuel rate. Thus, we model the interaction of the soil and the machine (soil-machine response) when preparing the seedbed with different machine configurations and environment states. For training and evaluating our model, we collected data from 16 field runs with varying soil types ranging from loamy sand to silty clay in Baden-Württemberg, Germany. Based on this data, we demonstrate that our novel model is able to predict the soil-machine response on previously unseen fields with high accuracy – in our setting with a relative root mean squared error (rRMSE) between 12.2 and 14.4 %. Our proposed model has two main advantages: (1) it allows farmers to plan the optimal machine configuration in advance by simulating different machine configurations and thus reduce the complexity for the operator and (2) it can serve as a basis for the training of intelligent control agents.
