Deep learning models for estimation of paddy rice leaf nitrogen concentration based on canopy hyperspectral data

Abstract

Rapid and nondestructive detection of crop nitrogen status is crucial to precision agriculture management. Hyperspectral remote sensing has been proposed to be a powerful tool for expediently monitoring crop nitrogen status. However, conventional regression methods and machine learning (ML) are difficult to utilize the full information of hyperspectral data, and deep neural networks (DNN) usually require a huge number of training data. Therefore, we attempt to construct deep learning models with a small amount of data and achieve accurate estimation of leaf nitrogen concentration (LNC). A two-year field experiment of paddy rice with four nitrogen levels were conducted at Jianli, Hubei province, China. A total of 216 samples containing canopy hyperspectral data and rice LNC were measured during two growing seasons. Based on the first derivative of hyperspectral data, a new deep learning model (deep forest, DF) was constructed for LNC estimation and compared with two traditional machine learning models (random forest, RF and support vector machine, SVM) and one deep neural network model (multi-layer perceptron, MLP). The results showed that, based on a small number of hyperspectral data, deep forest acquired higher accuracy than MLP. And the optimal estimation (R2 = 0.919, RMSE = 0.327) was obtained by the deep forest model based on full-wave band spectrum (350-2500 nm). Between two classical machine learning models, random forest achieved better results than SVM, but both methods were unstable. In conclusion, deep forest improved the prediction accuracy and model robustness for all band situations, and alleviated the degree of overfitting by multi-grained scanning. These results can provide a deep insight to detect crop nitrogen status rapidly when confronted with limited data.