A meta-analysis of crop leaf nitrogen, phosphorus and potassium content estimation based on hyperspectral and multispectral remote sensing techniques

Purpose

Real-time monitoring of essential nutrient status is crucial for improving fertilizer efficiency and enhancing crop productivity. Hyperspectral and multispectral remote sensing provide effective, non-invasive tools for estimating crop leaf nitrogen, phosphorus, and potassium content (LNC, LPC, and LKC). Therefore, a comprehensive evaluation of these technologies is needed.

Methods

We conducted a meta-analysis of studies from 2000 to 2023 to identify spectral bands for estimating LNC, LPC, and LKC. Subsequently, nutrient estimation models using Partial Least Squares Regression (PLSR), Random Forest (RF), and Support Vector Regression (SVR) were developed based on 4 years of oilseed rape field data, to verify identified the sensitive bands.

Results

The meta-analysis revealed an increasing research focus on nutrient estimation from 2017 to 2023, with wheat and rice as the primary crops investigated. Among the three nutrients, LNC was the most frequently analyzed. Commonly adopted modeling approaches included PLSR, Artificial Neural Networks (ANN), SVR, and RF. At the canopy level, LNC exhibited its most sensitive bands within 550–2030 nm, while at the leaf level, the sensitive range was 400–780 nm. LPC was responsive in 517–995 nm and 2030–2269 nm at the canopy level, while responsive in 545–995 nm and around 2166 nm at the leaf level. The bands sensitive to LKC were observed in 519–976 nm and 1513–2058 nm at the canopy level, and 545–995 nm at the leaf level. The RF model consistently achieved the highest prediction accuracy among models based on the identified sensitive bands. At the canopy level, LNC was estimated with the highest accuracy (R²=0.81, RMSE=0.39 %), followed by LPC (R²=0.75, RMSE=0.09 %) and LKC (R²=0.70, RMSE=0.34 %). At the leaf level, LNC again showed the best performance (R²=0.82, RMSE=0.37 %), followed by LKC (R²=0.74, RMSE=0.30 %) outperforming LPC (R²=0.66, RMSE=0.09 %).

Conclusions

This study provides a comprehensive evaluation of hyperspectral and multispectral technologies for crop nutrient estimation. The sensitive spectral bands and modeling approaches identified through meta-analysis enable accurate estimation of LNC, LPC, and LKC.