Identifying optimal ground feature classification and assessing leaf nitrogen status based on UAV multispectral images in an apple orchard

In the case of uneven vegetation coverage, it is still facing many problems that using UAV-remote sensing to assess the canopy nutrient status. The research objective is to determine the optimal ground feature classification and establish the leaf nitrogen concentration (LNC) inversion model.

UAV‑Phantom 4 multispectral platform was used to acquire apple orchard images. The ground features of remote sensing image were classified with the minimum distance, maximum likelihood, and object-oriented feature extraction classifications (MDC, MLC, OFEC), while spectral vegetation indices were used to perform LNC inversion using the backpropagation neural network (BP) and extreme learning machine (ELM). Further, genetic algorithm (GA) and particle swarm optimization (PSO) were used to optimize two inversion models.

Compared with the MDC, the overall accuracy of MLC increased 20.30% and 26.69% in 2021 and 2022, while the OFEC increased 30.48% and 31.04%, respectively. The LNC inversion model of BP and ELM produced an acceptable performance (R²_c > R²_p > 0.60, RRMSE < 15%). The use of GA and PSO algorithms did not improve the prediction performance of the BP inversion model. Compared with ELM, the RRMSE of GA_ELM and PSO_ELM inversion models decreased 2.63% and 20.39%. Furthermore, the PSO_ELM inversion model decreased the RRMSE 10.95% compared to the BP model.

The combination of reverse thinking iterative approach with the PSO algorithm significantly enhances the predictive performance of the ELM inversion model, which can present a rapid assessment method for leaf nitrogen nutrition diagnosis based on the relationship between LNC and fruit yield.