科研进展

【论文题目】 Machine learning models fed with optimized spectral indices to advance crop nitrogen monitoring

【论文题目】优化光谱指数助力机器学习算法提升作物氮素监测

【作者】 Haibo Yang, Hang Yin, Fei Li, Yuncai Hu, Kang Yu

杨海波，尹航，李斐，胡云才，于康

【摘要】

【Abstract】

Context: Remotely estimating plant nitrogen concentration (PNC) at the vegetative growth stage plays a crucial role in the precision N management of field crops. However, the great challenges still remain on how to overcome the impact of canopy structure variation and the ‘N-dilution’ effect on the accuracy of PNC assessment using spectral indices (SIs).

Objective: This study was to apply machine learning (ML) algorithms fed with the optimized spectral indices (OSI), sensitive spectral bands (SSB), and full-spectrum (FS) to improve the prediction accuracy of PNC in critical vegetative growth stages of wheat, maize, rice, potato and the across crops. Methods: Four ML algorithms including the partial least squares regression (PLSR), random forest (RF), support vector regression (SVR), and artificial neural network (ANN) were compared for their efficacies in predicting PNC from ten field trials in different locations from 2005 to 2016. Results: The input variable had a non-negligible influence on the performance of ML models. The OSI was the most efficient input variable for the tested ML algorithms in predicting PNC. The OSI-based RF models showed consistent outperformance compared to other models regardless of crops. The coefficient of determination (R²) was 0.51-0.85 and the root means square error (RMSE) was 0.15%-0.34% in the experimental validation datasets. By choosing PNC-related spectral features across crops, the OSI-based RF models increased prediction accuracy by 10%-31% compared with the best OSI-simple regression models, which was because the OSI-based RF models may be independent of canopy structure or the "N-dilution" impact. The simulated datasets based on the PROSAIL model and satellite multispectral bands further validated the results.

Conclusions: Our study concludes that the OSI-based RF is a robust and effective model to predict crop PNC at the vegetative growth stage.

Significance: The results of this study may provide insights into how to improve PNC assessment using the OSI-based RF models and deploy ML-based N recommendation models in the next generation of crop sensors.

【Keywords】Machine learning, Crops, Nitrogen, Remote monitoring