Corn variable-rate seeding decision based on gradient boosting decision tree model

Variable-rate seeding (VRS) technology can adjust the seeding rate according to the growth environment of the crop so as to improve crop yield and resource utilization. The information collection of sensor-based VRS is carried out at the same time as the seeding operation, which improves the operation efficiency and reduces the operation cost. Adjusting the seeding rate according to the indicators that characterize soil fertility is the key to sensor-based VRS technology. However, most of the current research on seeding rate decisions is based on delineating farmland management zones, which cannot be applied to sensor-based VRS. In most studies, traditional linear regression was used to establish the relationship among soil indicators, seeding rate, and yield, which did not consider the impact of weather and management factors on yield. Given the above problems, in this study, corn yield prediction model based on gradient boosting decision tree (GBDT) algorithm was established by combining soil organic matter (SOM) data, weather data and management data, and was compared with the model built by random forest (RF) algorithm. Innovative seeding rate decision rules were developed by using the GBDT corn yield prediction model to simulate corn yield responses to a series of SOM contents and seeding rates. The result showed that the GBDT model (R2cv = 0.799) was better than the RF model (R2cv = 0.749). The simulation results of the GBDT model indicated that the yield showed a parabolic form with the increase of seeding rate under the same SOM, the yield increased first and then decreased with the increase of SOM under the same seeding rate, there was an agronomic optimum seeding rate (AOSR) under each SOM and the AOSR increased in a stepped shape with the increase of SOM. The following decision rules were obtained by summarizing the simulation results: when 10 g/kg ≤ SOM ≤ 12 g/kg, seeding rate = 84,900 seeds/ha; when 12 g/kg < SOM ≤ 13 g/kg, 84,900 seeds/ha < seeding rate ≤ 90,600 seeds/ha; when 13 g/kg < SOM ≤ 14 g/kg, 90,600 seeds/ha < seeding rate ≤ 92,110 seeds/ha; when 14 g/kg < SOM ≤ 20 g/kg, 92,110 seeds/ha < seeding rate ≤ 93,000 seeds/ha; when 20 g/kg < SOM ≤ 26 g/kg, 93,000 seeds/ha < seeding rate ≤ 93,880 seeds/ha. The crop yield prediction model is established through a machine-learning algorithm to create a controllable environment for exploring the change of yield with soil attributes and seeding rate, which is a promising method to study sensor-based variable-rate seeding decisions.