Document Type : Original Research Paper
Authors
Department of Surveying Engineering, Faculty of Civil, Water and Environmental Engineering, Shahid Beheshti University, Tehran, Iran
Abstract
Background and Objectives: Nowadays, getting land cover and land use information is crucial due to the growing number of uses for this data. The primary method for obtaining this information is considered to be through the utilization of remote sensing images. Image classification techniques should be employed so as to extract land cover and use from these images. Deep learning techniques can be utilized effectively to the classification of land cover and land use simply because of their great potential in image classification. But there are also challenges when applying these techniques as well. Model overfitting is one of the most common issues when utilizing deep learning algorithms. Another major issue with these methods is that they demand a significant amount of data during the training stage. Additionally, gradient exploding/vanishing and determining the suitable architecture are further challenges associated with these methods for extracting land cover and use from remote sensing imagery.
Methods: The main objective of this research is to employ different techniques to overcome the challenges to achieve high classification accuracy. To solve the problem of model overfitting, dropout and early stopping approaches were utilized to ensure that the accuracy of the training and test data were close. The data augmentation strategy can prevent model overfitting in addition to addressing the lack of training data. As a result, this method was employed to augment training data and also avoiding model overfitting. The gradient clipping strategy was additionally used in this study to mitigate gradient exploding and vanishings in deep learning models. This study used the ResNet18 model to classify the EuroSat dataset, enabling us to obtain highly effective classification accuracy.
Findings: Initially, this architecture was used with with the early stopping strategy, and the model had an overall accuracy of 91.19 percent and a kappa coefficient of 0.9018. The data augmentation technique was then applied to the same model, and the model achieved an overall accuracy of 91.78 percent with a kappa coefficient of 0.9085, surpassing the previous stage. In the last stage, a dropout method with a rate of 0.5 and a gradient clipping with a threshold of 0.1 were added to the previous model, and the model achieved an overall accuracy of 93.11 percent and a kappa coefficient of 0.9233, which was more accurate than the previous two stages.
Conclusion: These results indicate that the EuroSat's land cover and land use classification accuracy in the final stage was higher than in prior stages.
Keywords
Main Subjects
COPYRIGHTS
© 2025 The Author(s). This is an open-access article distributed under the terms and conditions of the Creative Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
)