This paper presents a supervised machine learning (ML)-based protection approach for identifying different types of transmission lines' faults, including faults during power swing, at several locations and variable fault resistances. Descriptive statistical features used in classification were extracted from spectrograms of measured signals at the relay point. Decision trees (DT)., support vector machines (SVM), k-nearest neighbors (k-NN), boosted and bagged ensemble tree classifiers were used for classification. Four feature selection algorithms, namely neighborhood component analysis (NCA), minimum redundancy maximum relevance (mRMR), sequential feature selection (SFS), and fit ensemble of learners, were applied to reduce the number of features. The synthetic minority class oversampling technique (SMOTE) balances the data to prevent biases toward the majority (non-fault) class. The results show that ensemble bagged trees with SMOTE achieve the maximum accuracy and the minimum false-negative rates. Feature selection algorithms show no improvement in the performance of the classifiers. The best classifier was then tested using data from an unseen scenario and showed accurate detection of the fault events. The classifier, however, was not able to detect faults during power swing after integrating a photovoltaic (PV) plant behind the relay point in the protected line.