Hyperspectral remote sensing data have great potential for plant classification, monitoring, and mapping due to their considerable spectral information. A spectral library can be used for the automatic interpretation of remote sensing data, which is an efficient tool for the classification of plant species. Given that similar spectral signatures are usually found among different plant species, it is critical to identify plant-sensitive spectral features when constructing a robust plant spectral library. This study proposed an approach to the establishment and application of a plant spectral library based on spectral data from ground objects. It included a spectral feature screening method for plant spectral response and a feature robustness analysis method suitable for monitoring, while unmanned aerial vehicle (UAV) hyperspectral imaging data was used to validate mapping results. This study was divided into an urban scenario and an agricultural scenario based on plant composition, and a set of spectral feature sensitivity methods based on Iterative Self-Organizing Data Analysis (ISODATA) and Jeffries–Matusita (JM) distance was proposed. At the same time, a robust spectral analysis method was developed based on illumination and noise disturbance. The performance of K-nearest neighbour algorithm (KNN), Random Forest (RF), and Support Vector Machine coupled with a Genetic Algorithm (GA-SVM) for plant classification modelling were compared. Results showed that the spectral indices obtained by sensitivity and robustness screening were effective for plant classification, and the GA-SVM model had the highest accuracy with overall accuracy (OAA) = 0.98, Kappa = 0.98 in the urban scenario, and OAA = 0.97, Kappa = 0.97 in the agricultural scenario. In addition, pixel-based crop classification validation based on UAV hyperspectral imaging also had high accuracy, and the plot level classification results were in good agreement with field survey results. Therefore, it is feasible to use a plant spectral library to assist in monitoring and mapping of plant species using hyperspectral remote sensing images at large scales. The effects of plant growth status, growth stage, and other changes on classification should be further studied.

中文翻译：

---

基于优化特征选择方法的植物谱库构建

高光谱遥感数据由于其大量的光谱信息，在植物分类、监测和制图方面具有巨大的潜力。光谱库可用于遥感数据的自动解译，是植物物种分类的有效工具。鉴于通常在不同植物物种中发现类似的光谱特征，在构建强大的植物光谱库时识别植物敏感的光谱特征至关重要。本研究提出了一种基于地物光谱数据的植物光谱库的建立和应用方法。它包括植物光谱响应的光谱特征筛选方法和适用于监测的特征稳健性分析方法，而无人机（UAV）高光谱成像数据用于验证映射结果。本研究根据植物成分分为城市场景和农业场景，提出了一套基于迭代自组织数据分析（ISODATA）和Jeffries-Matusita（JM）距离的光谱特征敏感性方法。同时，开发了一种基于光照和噪声干扰的鲁棒光谱分析方法。比较了 K-近邻算法 (KNN)、随机森林 (RF) 和支持向量机结合遗传算法 (GA-SVM) 对植物分类建模的性能。结果表明，通过敏感性和稳健性筛选获得的光谱指标对植物分类有效，GA-SVM 模型的准确率最高，在城市场景中总体准确度 (OAA) = 0.98，Kappa = 0.98，在农业场景中 OAA = 0.97，Kappa = 0.97。此外，基于无人机高光谱成像的基于像素的作物分类验证也具有较高的准确性，地块级分类结果与实地调查结果吻合良好。因此，利用植物光谱库辅助大尺度高光谱遥感影像对植物物种进行监测和制图是可行的。应进一步研究植物生长状态、生长阶段和其他变化对分类的影响。基于无人机高光谱成像的基于像素的作物分类验证也具有较高的准确性，地块级分类结果与实地调查结果吻合良好。因此，利用植物光谱库辅助大尺度高光谱遥感影像对植物物种进行监测和制图是可行的。应进一步研究植物生长状态、生长阶段和其他变化对分类的影响。基于无人机高光谱成像的基于像素的作物分类验证也具有较高的准确性，地块级分类结果与实地调查结果吻合良好。因此，利用植物光谱库辅助大尺度高光谱遥感影像对植物物种进行监测和制图是可行的。应进一步研究植物生长状态、生长阶段和其他变化对分类的影响。