USING GIS, EUROSAT DATASET, AND CONVOLUTIONAL NEURAL NETWORKS IN RESIDENTIAL LAND DETECTION IN THUA THIEN HUE PROVINCE, VIETNAM

Abstract

This paper evaluates the level of suitability of applying GIS, EuroSAT dataset, and Convolutional Neural Networks to recognize residential land in Thua Thien Hue, Vietnam. We employed EuroSAT dataset to train and validate the ResNet152 model. Then, the assessed model was used to detect residential land in the location. The classification accuracy assessment of the Resnet152 model trained with EuroSAT dataset is based on some metrics, including Overall Accuracy, Recall, Precision and F-score, and the residential land detection accuracy of trained Resnet152 was evaluated by using Overall Accuracy. The results show that the performance of Resnet152 over EuroSAT dataset acquires high quality, with the model's classification accuracy reaching 0.98 and all other metrics greater than 0.9. Besides, the model’s detection accuracy of the Resnet152 model and the ground-truth data is as high as 0.88. The most incorrect detection occurs between residential land and two other lands, namely industrial land and herbaceous vegetation. Finally, it is undeniable that this EuroSAT-based ResNet152 model has significant potential to apply in the central region of Vietnam.