Information on tree species across various spatial scales, from an individual tree to a forest stand and the broader landscape, contributes to an accurate and thorough understanding of forest conditions either as an individual characteristic or as an input of species-dependent models. However, tree species recognition is one of the most challenging tasks in forest remote sensing studies, due to the complexity of species compositions and canopy structures of forests, e.g., both cross-species similarities and intra-species variations commonly exist in spectral-, texture-, and structure- domains. Over the past two decades, the interest in using close-range sensing for tree species recognition has been rapidly growing. Recent research has highlighted the needs to further develop species recognition methods to elevate their performance in comparison with established remote sensing approaches, and to address new questions arising from spatial resolutions, data coverages, viewing geometries, and other data characteristics. This work provides an overview of the state-of-the-art of tree species recognition from close-range sensing data. The work summarizes the research works in the past decade, reviews the state of research, discusses prominent challenges, investigates impact factors, research gaps, and new potentials. Specifically, data from various sources, the features derived from each type of data, methodologies applied, and the targeted species are reviewed in detail. Relevant machine learning (ML) approaches are grouped into conventional ML and deep-learning (DL) categories. In each category, the reported studies/results are reviewed with respect to the spectral, spatial, and temporal domains of the used data sources, e.g., sensor and platform. Despite significant efforts in the field, the issues of automation, reliability, and robustness of the algorithms have only been partially resolved. The crucial elements in algorithm design what this work found and is worth careful consideration include forest types and stand conditions, seasonal variability and phenology, data characteristics and the corresponding feature selections, and the methodology. Future studies are recommended to focus on the fusion of multi-source data including passive and active multispectral data to integrate the spectral and structural information, the use of time-series data to enhance the role of phenological variances in species recognition, and the development of unsupervised DL techniques to improve the recognition accuracy and efficiency. It is also crucial to promote data sharing and open standards to facilitate international cooperation and communication.

中文翻译：

---

通过近距离传感进行树种识别：综述


不同空间尺度上的树种信息，从单棵树到林分以及更广泛的景观，有助于准确、透彻地了解森林状况，无论是作为个体特征还是作为依赖于物种的模型的输入。然而，由于森林物种组成和冠层结构的复杂性，树种识别是森林遥感研究中最具挑战性的任务之一，例如，光谱、纹理中普遍存在种间相似性和种内差异。 - 和结构域。在过去的二十年中，人们对使用近距离传感进行树种识别的兴趣迅速增长。最近的研究强调需要进一步开发物种识别方法，以提高其与现有遥感方法相比的性能，并解决由空间分辨率、数据覆盖范围、观察几何形状和其他数据特征引起的新问题。这项工作概述了近距离传感数据树种识别的最新技术。该著作总结了过去十年的研究工作，回顾了研究现状，讨论了突出的挑战，调查了影响因素、研究差距和新的潜力。具体来说，详细审查了来自各种来源的数据、每种类型数据的特征、所应用的方法以及目标物种。相关机器学习 (ML) 方法分为传统 ML 和深度学习 (DL) 类别。在每个类别中，所报告的研究/结果都根据所使用的数据源（例如传感器和平台）的光谱、空间和时间域进行审查。 尽管该领域付出了巨大努力，但算法的自动化、可靠性和鲁棒性问题仅得到部分解决。这项工作发现并值得仔细考虑的算法设计的关键要素包括森林类型和林分条件、季节变化和物候、数据特征和相应的特征选择以及方法。建议未来的研究重点关注包括被动和主动多光谱数据在内的多源数据的融合，以整合光谱和结构信息，利用时间序列数据增强物候差异在物种识别中的作用，以及开发无监督深度学习技术提高识别精度和效率。促进数据共享和开放标准以促进国际合作和交流也至关重要。