Enhancing grassland nitrogen estimation : a multiscale approach through optical reflectance spectroscopy and hybrid modeling techniques : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Earth Science at Massey University, Palmerston North, New Zealand

Abstract

Optical remote sensing technology has emerged as a powerful tool for assessing vegetation characteristics, particularly nitrogen (N) concentration (N%) in heterogeneous grasslands. Accurate estimation of N% is crucial for farmers, as it directly influences grassland productivity and plays a key role in sustainable land management. Accurate N assessments optimize fertilizer use, boosting productivity, lowering costs, and enhancing environmental modeling to address impacts such as N leaching and greenhouse gas emissions. Despite significant progress, challenges and knowledge gaps remain, highlighting the need for continued research to fully harness remote sensing’s potential in agricultural management and its impact on livestock productivity. This thesis aims to advance the accurate estimation of grassland N% by integrating physically-based, empirical-statistical, and hybrid models using optical reflectance spectroscopy data. The research focuses on three primary objectives:

1. To estimate N% in grasslands using optical reflectance spectroscopy, data will be collected across multiple scales, including ground-, leaf-, canopy-, and satellite-scale observations.

2. To improve the universality and adaptability of grassland N% models through a hybrid approach that combines data from various optical sensors across multiple scales.

3. To account for and quantify uncertainties in grassland N% prediction models.

The thesis addresses the challenge of uncertainty by conducting a comprehensive analysis of its sources and developing methods, such as Physically Informed Neural Networks (PINN), to account for them. Key strategies include data fusion techniques for integrating diverse data sources and improving atmospheric correction methods. A unified methodology combining empirical-statistical and physically-based approaches is proposed to enhance generalization. Machine learning algorithms play a pivotal role in feature selection and optimization, further improving model accuracy and transferability. The developed methods undergo evaluation using independent validation data collected from heterogeneous grasslands across different periods and locations. Results demonstrate that integrating physically-based and empirical-statistical approaches significantly improves model accuracy and transferability, providing a deeper understanding of the factors influencing vegetation traits. This thesis highlights the importance of advanced techniques, including machine learning, deep learning algorithms, Radiative Transfer Models (RTM), and data fusion methods, for precisely characterizing vegetation traits, contributing to more sustainable and efficient grassland management practices.