Comparative analysis of the effects of different dimensionality reduction algorithms on hyperspectral estimation of total nitrogen content in wheat soils

Context

The level of soil nitrogen supply profoundly impacts the growth, development, and yield formation capacity of winter wheat. Excessive use of nitrogen-based fertilizers in current agricultural practices has negative consequences on both the environment and crop growth. Therefore, real-time, non-destructive estimation of soil total nitrogen content using hyperspectral remote sensing technology is crucial for advancing crop fertilization strategies and precision agriculture.

Objectives

(1) Explores the effects of different dimensionality reduction algorithms on hyperspectral estimation of soil total nitrogen content in wheat fields. (2) Investigates the optimal model for hyperspectral detection of total nitrogen content in wheat field soils in the Jinzhong region of Shanxi Province.

Methods

This study integrates various preprocessing methods and applies four dimensionality reduction algorithms—principal component analysis (PCA), singular value decomposition (SVD), unrelated variable elimination (UVE) and random forest (RF)—to reduce the data dimensions. Support vector regression (SVR) and back propagation neural network (BPNN) models for estimating soil total nitrogen content were then constructed and compared with gradient boosted decision tree (GBDT).

Results

The feature extraction algorithms PCA and SVD produced the same principal components and cumulative contributions when reducing the dimensionality of hyperspectral data. The number of characteristic bands selected by UVE was much smaller than that selected by RF. The characteristic bands selected by RF spanned the visible, near-infrared, and mid-infrared wavelength ranges, while those selected by UVE were mostly located within the visible light wavelength range. The modelling results following PCA and SVD dimensionality reduction were relatively similar, while the models based on RF-selected bands showed little change compared to full-spectrum band modeling. The SVR model constructed using multiplicative scatter correction (MSC) preprocessing and SVD dimensionality reduction had the highest accuracy in estimating soil total nitrogen content. (R_c²=0.87, R_v²=0.85; RMSE_c=0.13, RMSE_v=0.14; RPD_c=2.82, RPD_v=2.55; MAE_c=0.10, MAE_v=0.10)

Conclusions

Dimensionality reduction algorithms significantly contribute to the development of hyperspectral estimation models for soil total nitrogen content. The feature extraction algorithm (PCA and SVD) shows more obvious effect in improving the spectral modeling accuracy compared to the feature selection algorithm (UVE and RF). The optimal estimation model combination for hyperspectral detection of total nitrogen content in wheat field soils is MSC+SVD+SVR.