Detection and Mapping of Slow-moving Landslides with Pleiades-1 Data

Abstract

 The study utilised the k-NN and SVM algorithms, Pleiades-1 data, and object-based image analysis for the classification and mapping of slow-moving landslides in Kutlugün in Maçka district of Turkey. The approaches employed were based on the investigation of the influences of the training sample sizes and type (balanced/imbalanced) on the accuracy of the classification results. A total of 128 and 134 small landslides were detected using k-NN and SVM algorithms, respectively. The SVM method had higher producer accuracy (85.9%), user accuracy (89.4%) and kappa index (0.82) compared to the k-NN algorithm that had producer accuracy (83.1%), user accuracy (86.0%) and kappa index (0.80). Using the imbalanced datasets, the classification accuracy of SVM was not significantly different among the five different training sample sizes, with 83.75% for the lowest dataset (dat2) and 85.12% for the highest dataset (dat5). On the other hand, the classification accuracy of k-NN was significantly different between the smallest sample size (dat1, 79.80%) and the largest sample size (dat5, 84.59%). When balanced datasets were used, the SVM algorithm produced accuracy of 85.07% compared to k-NN classifier (84.10%), showing that there was no pronounced difference on the performance of the two classifiers on different training sample sizes. The overall evaluation of the two algorithms show that using supervised classification at object level with Pleiades-1 image, for slow-moving landslide detection and mapping, is possible and can be improved.