# Inversion of kiwifruit canopy nitrogen using UAV multispectral technology and ensemble learning

## TL;DR

This study uses drones and machine learning to accurately measure nitrogen levels in kiwifruit canopies in mountainous regions.

## Contribution

The study introduces an ensemble learning approach combined with UAV multispectral data for precise nitrogen inversion in vine canopies.

## Key findings

- Eleven spectral variables showed significant correlation with nitrogen content.
- The Boosting ensemble model achieved high inversion accuracy (R2 = 0.89, RMSE = 0.50, RPD = 2.99).
- A heatmap effectively visualized the spatial distribution of nitrogen in the orchard.

## Abstract

Accurate nitrogen monitoring is a key prerequisite for the high-quality, high-yield, and sustainable cultivation of mountainous kiwifruit, yet the complex topography of mountainous regions and the unique vine canopy structure of kiwifruit limit the applicability of traditional monitoring methods. In this study, a low-latitude, high-altitude mountainous kiwifruit orchard in Yunnan, China, was selected as the study area, with a focus on the fruit expansion stage (August). We integrated UAV multispectral technology and ensemble learning algorithms to perform canopy nitrogen inversion. Canopy images were acquired via UAV, and 278 experimental plots were synchronized with ground-based measured nitrogen data collection. Twenty-five spectral variables, five single models, and three ensemble learning models were constructed, and the SHAP method was employed to analyze the contribution of each spectral feature to nitrogen inversion. The results showed that eleven spectral variables were significantly correlated with nitrogen content, PLSR was the best single model, and the Boosting ensemble model had the best inversion accuracy (R2 = 0.89, RMSE = 0.50, RPD = 2.99). The heatmap generated by the model clearly depicts the spatial distribution of nitrogen. This study confirms the feasibility of coupling UAV multispectral technology with ensemble learning algorithms for nitrogen inversion, optimizes the nitrogen inversion technical system for mountainous vine fruits, and provides theoretical and technical support for the popularization of smart agriculture in mountainous orchards.