# Construction and UAV-based inversion of integrated nitrogen diagnosis index for cotton using multispectral imagery

## TL;DR

This study develops a new index for diagnosing cotton nitrogen status using drone-based multispectral imagery and machine learning, improving precision fertilization in drip-irrigated systems.

## Contribution

The novel Integrated Nitrogen Diagnosis Index (INDI) and its inversion using XGBoost provide a stable and accurate method for nitrogen diagnosis across cotton growth stages.

## Key findings

- XGBoost achieved the highest inversion accuracy for INDI with R2 = 0.85 and RMSE = 0.61.
- INDI showed strong correlation with NNI across growth stages, with the highest accuracy at the boll-setting stage.
- INDI spatial distribution maps effectively distinguished nitrogen differences under varying water-nitrogen treatments.

## Abstract

Accurate and stable diagnosis of cotton nitrogen status across growth stages is essential for precision fertilization in drip-irrigated systems. However, the instability of conventional nitrogen-related indicators across different phenological stages often reduces diagnostic performance and limits their broader application.

A field experiment was conducted in Xinjiang, China, under four irrigation levels (60%, 80%, 100%, and 120% ETc) and four nitrogen application rates (0, 245, 300, and 350 kg N ha-1). UAV multispectral imagery was acquired at the squaring, flowering, boll-setting, and boll-opening stages. Based on ground-measured leaf area index (LAI) and upper-canopy leaf nitrogen weight (LNWupper), an Integrated Nitrogen Diagnosis Index (INDI) was developed. Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting (XGBoost) models were used to evaluate the inversion performance of INDI. In addition, the nitrogen nutrition index (NNI), derived from the critical nitrogen dilution curve, was used to validate the diagnostic stability of INDI.

Multispectral vegetation indices were strongly correlated with LAI, LNWupper, and INDI, with red-edge- and near-infrared-based indices showing the highest sensitivity. Among the three models, XGBoost achieved the best inversion accuracy for INDI (R2 = 0.85, RMSE = 0.61). INDI was significantly correlated with NNI across growth stages, with R2 values of 0.58, 0.77, 0.81, and 0.70 at the squaring, flowering, boll-setting, and boll-opening stages, respectively, and the highest accuracy observed at the boll-setting stage. Moreover, the spatial distribution maps of INDI effectively distinguished nitrogen differences under different water–nitrogen treatments and were consistent with NNI-based classifications.

INDI accurately captured nitrogen dynamics throughout cotton growth, and the INDI–XGBoost framework provided a robust approach for high-precision spatial nitrogen diagnosis. These results support precision fertilization management in drip-irrigated cotton fields in Xinjiang.