Automatic training data extraction approaches are new solutions for supervised classification. Most of the researches are aimed to develop automatic methods since conventional methods are time- and cost-consuming. Also, in conventional methods the results are impacted by the expertise of operators. On the other hand, factors such as size, number, distribution and homogeneity of training samples affect the overall accuracy. Therefore, examining the impact of these factors is time-consuming in traditional methods. Newly presented approaches have removed the problems of traditional methods although they have their own issues such as more detailed data in hyper-spectral images or heterogeneity inner different classes. These factors influence the accuracy of classification. In this paper, we study the behavior of raw data and extracted features separately, and then identify a stable pattern for automatic recognition of each class. The study is based on frequency distribution and statistical analysis of spectral behavior of ground features. Investigation on the results exposes that some classes have stable patterns and can easily be identified according to their behavior. The behavior of each class depends on selected bands and features, and can be regular or irregular. For instance, water class has low density distribution that makes it difficult to fit a distribution function, but it is distinguishable thorough unpredictable distribution function. In the other hand, urban class has regular pattern and an appropriate distribution can be fitted to the frequency of the features extracted from its sample dataset. Although an individual class may be recognized through several features simultaneously, using whole available features is time-consuming. So, just intrinsic combination of features should be used for classification. In the research, the behaviors of studied classes have been presented in different charts and diagrams, and ultimately, a new algorithm (in flowchart format) has been presented for intelligent sampling and training data extraction based on statistical analysis of spectral behavior of classes.

中文翻译：

---

开发一种基于特征光谱行为统计分析的新算法，用于自动提取训练数据以对遥感图像进行分类

自动训练数据提取方法是监督分类的新解决方案。大多数研究旨在开发自动化方法，因为传统方法既费时又费钱。此外，在传统方法中，结果受操作员专业知识的影响。另一方面，训练样本的大小、数量、分布和同质性等因素会影响整体准确率。因此，在传统方法中检查这些因素的影响是耗时的。新提出的方法消除了传统方法的问题，尽管它们有自己的问题，例如高光谱图像中的数据更详细或不同类别内部的异质性。这些因素影响分类的准确性。在本文中，我们分别研究原始数据和提取特征的行为，然后确定一个稳定的模式用于自动识别每个类。该研究基于频率分布和地物光谱行为的统计分析。对结果的调查表明，某些类具有稳定的模式，可以根据其行为轻松识别。每个类的行为取决于选定的波段和特征，可以是规则的或不规则的。例如，水类具有低密度分布，难以拟合分布函数，但可区分透彻的不可预测的分布函数。另一方面，城市类具有规则的模式，可以根据从其样本数据集中提取的特征的频率来拟合适当的分布。尽管可以同时通过多个特征识别单个类，但使用整个可用特征是耗时的。因此，应该仅使用特征的内在组合进行分类。在研究中，研究类的行为以不同的图表和图表的形式呈现，最终提出了一种基于类的光谱行为统计分析的智能采样和训练数据提取的新算法（流程图格式）。