Investigating the performance of pixel-based and object-basedimage classification methods

In this research, large-scale remotely sensed datasets namely SENTINEL-2 and LANDSAT- 9 were classified by five supervised algorithms, Support Vector Machine (SVM), Maximum Likelihood Classification (MLC), Minimum Distance Classification (MDC), Artificial Neural Net Classification ( ANN) and eCognition - Support Vector Machine (ECO-SVM) depending on the amount of training samples. The overall accuracy ranged from 200 to 403 sample sizes. In the study, it was observed that large training sets gave higher accuracy rates compared to small training sets. For both SENTINEL-2 and LANDSAT-9 data, higher accuracy rates were obtained when using large training sets. Accuracy varied between classifications with different training samples, even when using the same classifier and training sample size.Therefore, depending on the size of the training set, some algorithms may be more accurate than others.According to the results of the study, Results of object-based classifications have provided higher accuracy rates than pixel-based classification. SVM provided the highest accuracy in both SENTINEL-2 and LANDSAT-9 data. In SENTINEL-2, the pixel-based classification algorithm ANN gave the second highest result, while in LANDSAT-9, the pixel-based SVM gave the second highest result. In general, higher accuracy rates were obtained when using larger sample sizes.