Impact of training data size on classifiers when coarse resolution imageries were used for regional land cover mapping

This work analyzes the impacts of training sample size on the performance of supervised classification methods when coarse resolution imageries are employed for regional land cover mapping. We utilized FegnYun-3C composite imageries with 1km spatial resolution and random forest (RF) and support vector machine (SVM) algorithms that were trained and tested with five sets of reference datasets: 66/34, 69/31, 73/27, 76/24 and 79/21. The results show that the performance of the two algorithms increases with increasing the size of the training examples until a certain point, and achieves the maximum accuracy (0.86 for RF and 0.84 for SVM) when the ratio was 76/24. However, considering the 79/21 (train/test) ratio made no change in accuracy, implying increasing a training dataset beyond a certain limit has no effect. Moreover, despite the size of training samples employed, the RF outperformed the SVM in contrary to the claim that SVM yields a better accuracy in case of scarce training data by pervious studies.