Abstract Electric Distribution Transformers (EDTs) are used in many areas for converting the voltage used in the transmission lines to a suitable voltage. One of the specific use areas of EDTs is to provide irrigation of agricultural areas. However, there is a great difficulty in detection of unregistered EDTs in these fields. The illegal usage of them damages the electric distribution companies and cause serious damages by disrupting the electricity grid system. To overcome these problems, Sentinel-1 SAR images give good results when used together with machine learning approaches. This work was carried out on agricultural lands in Diyarbakir province, Turkey and it consists of several stages. In first place, GPS coordinates of 178 sample points were recorded and 32 Sentinel-1 SAR images were obtained between March-May & 2016–2018. In the second phase, backscattering coefficients in different polarizations were generated for each point in these dates. The features obtained from 32 Sentinel-1 SAR images were then added successively and a feature vector with a total length of 64 was produced for each sample point. After that, different machine learning techniques [Extreme Learning Machine (ELM), K-Nearest Neighbor (K-NN), Naive Bayes Algorithm (NBA) and Support Vector Machine (SVM)] were applied to the obtained datasets and ELM was determined as best machine learning approach. A machine learning based feature selection (ELM-FS) method was then applied to all feature vectors and the most relevant features were determined according to their impact values. In the last stage, feature sets obtained from ELM-FS method were used separately as inputs of ELM technique and the unregistered EDTs were successfully detected with 85.47% mean accuracy.

中文翻译：

---

通过机器学习方法借助 Sentinel-1 SAR 图像检测农田中未注册的配电变压器

摘要 配电变压器 (EDT) 用于许多领域，用于将传输线中使用的电压转换为合适的电压。EDT 的具体使用领域之一是为农业区提供灌溉。然而，在这些领域检测未注册的 EDT 存在很大的困难。非法使用它们对配电公司造成损害，并通过扰乱电网系统造成严重损失。为了克服这些问题，Sentinel-1 SAR 图像与机器学习方法一起使用时会产生良好的效果。这项工作是在土耳其迪亚巴克尔省的农田上进行的，它包括几个阶段。首先，记录了 178 个样本点的 GPS 坐标，并在 3 月至 5 月和 2016 年至 2018 年期间获得了 32 张 Sentinel-1 SAR 图像。在第二阶段，为这些日期中的每个点生成不同极化的后向散射系数。然后依次添加从 32 张 Sentinel-1 SAR 图像中获得的特征，并为每个样本点生成一个总长度为 64 的特征向量。之后，将不同的机器学习技术[极端学习机 (ELM)、K-最近邻 (K-NN)、朴素贝叶斯算法 (NBA) 和支持向量机 (SVM)] 应用于获得的数据集，ELM 被确定为最好的机器学习方法。然后将基于机器学习的特征选择 (ELM-FS) 方法应用于所有特征向量，并根据其影响值确定最相关的特征。在最后阶段，