土壤养分准确预测对于精准科学的控制施肥,从而降低生产成本提升玉米产量具有非常重要的作用,本文提出一种基于多尺度局部增强Transformer的土壤养分预测方法。首先,针对传统玉米地块土壤数据采样不考虑时间问题,设计构建了基于时间的土壤数据样本数据库,并针对其存在的时间尺度退化问题,引入多尺度局部增强模块用于强化Transformer网络的局部特征提取,解决土壤养分数据存在的时间依赖导致无法实现全局学习的问题；其次,针对传统卷积存在的数据泄漏现象,设计了考虑时间序列的序列卷积操作,要求卷积核只可采用数据库中的历史数据进行学习训练,避免其学习模型过分依赖将来数据问题,能访问过去的数据点,从而确保模型在建模时间序列时不会依赖未来数据,提升了模型学习精度和合理性。最后,通过消融实验验证了所提模型在内蒙古某玉米地块土壤养分预测上的准确性,为玉米地块的低成本土壤养分快速检测和精准科学施肥提供了科学依据。

Accurate prediction of soil nutrients plays a very important role in precise and scientific control of fertilization, thereby reducing production costs and increasing corn yield. This paper proposes a soil nutrient prediction method based on multi-scale local enhanced Transformer. Firstly, to address the issue of traditional corn plot soil data sampling not considering time, a time-based soil data sample database was designed and constructed. In response to the problem of time scale degradation, a multi-scale local enhancement module was introduced to enhance the local feature extraction of the Transformer network, solving the problem of the inability to achieve global learning due to the time dependence of soil nutrient data; Secondly, in response to the data leakage phenomenon of traditional convolution, a sequence convolution operation considering time series is designed, which requires the convolution kernel to only use historical data in the database for learning and training, avoiding the problem of the learning model overly relying on future data. It can access past data points to ensure that the model does not rely on future data when modeling time series, thereby improving the accuracy and rationality of model learning. Finally, the accuracy of the proposed model in predicting soil nutrients in a corn plot in Heilongjiang Province was verified through ablation experiments, providing a scientific basis for low-cost rapid detection of soil nutrients and precise scientific fertilization in corn plots.

S513.062；TP391

内蒙古自治区十四五规划课题 (NGJGH2021459)