Convolutional Neural Network for Thailand's Eastern Economic Corridor (EEC) land cover classification using overlapping process on satellite images

Land cover is a powerful tool in urban management as a source of information to support authorities' decision making. In this paper, land cover of Eastern Economic Corridor cities in Thailand is performed using the aggregation of results from two Convolutional Neural Network models, with both using the same architecture. The first model is for the detection of water and the second is for the classification of land type consisting of 3 classes: city, forest and crop. In Firstly, the size of the 4 class existing dataset is increased resulting in an accuracy of 98% and 93% for the binary and three class CNN model respectively. To improve the land cover on satellite images an overlapping process is introduced in order to reduce the classification area from 0.25 km(2) to 0.004 km(2), using the same image resolution of 8 m per pixel. The use of the overlapping allows to propose a largely better land cover where the contour of the detected class is well produced. Moreover, this methods shows its better ability to detect smaller surface size and especially for the water, crops and forest class. However, it is showed that the overlapping method does not improved the accuracy of the prediction that is mainly related to the dataset size. Finally, the robustness of the proposed method to quickly perform a global land cover with limited computer power is demonstrated.