Detecting Vegetation Encroachment Along Power Lines with Handcrafted Features and an RBF-SVM Classifier

Abstract

Vegetation encroachment around overhead power lines and substations is a leading cause of unplanned outages, equipment damage, and wildfire ignition. Manual right-of-way (ROW) inspection is slow, expensive, and error-prone, which motivates automated remote-sensing solutions. While deep con-volutional networks dominate modern image classification, they require large annotated corpora and substantial computation, limiting their use on edge and embedded inspection platforms. This study presents an interpretable pipeline that uses a Support Vector Machine (SVM) with a radial basis function (RBF) kernel and manually created color and texture descriptors to classify satellite/aerial picture patches into high- and low-density vegetation. From each patch we extract per-channel colour moments (mean, standard deviation, skewness, kurtosis), Gray- Level Co-occurrence Matrix (GLCM) descriptors, and a Local Binary Pattern (LBP) texture summary. Evaluated on a balanced corpus of 12,351 image patches, the model attains a test accuracy of 98.2%, a macro-averaged F1 score of 0.983, and an Receiver Operating Characteristic - Area Under the Curve (ROC–AUC) of 0.998. Grouped 5-fold cross-validation yields a mean macro- F1 of 0.983, confirming that the performance is stable and not an artefact of a single split. Permutation feature-importance analysis shows that colour moments—particularly the red and green channel means—are the dominant discriminative cues. The proposed approach is accurate, reproducible, computationally inexpensive, and interpretable, making it well suited for real-time, resource-constrained infrastructure-monitoring deployments.