LAND COVER AND LAND USE CLASSIFICATION IN REMOTE SENSING DATA USING DEEP LEARNING

Abstract

We compared the results from the state-of-the-art studies and other built models with those from the UCM, SIRI-WHU, and RSSCN7 datasets after building and training the DL models on those datasets. The first experiment was designing CNN-FE. Results in this experiment showed performance improvements for the CNN-FE model when compared to state-of-the-art baseline studies and the VGG-19 pretrained model. Additionally, the CNN-FE model's performance was better when trained on the UCM dataset compared to when trained on the SIRI-WHU dataset. The second experiment was building the TL model. We used the InceptionV3, Resnet50V2, and VGG19 pretrained models for LCLU classification in the UCM dataset, with the TL model trained with bottleneck feature extraction. Based on these experiments, the TL model was developed, with improved results of 92.46, 94.38, and 99.64 in Resnet50V2, InceptionV3, and VGG19, respectively. The third experiment was the comparative evaluation of the CNN-FE and TL with fine-tuning. In this experiment, the fine-tuning model has outperformed the CNN-FE and TL in both the UCM and SIRI-WHU datasets. The fourth experiment was building the DL models, such as EfficientNetB7, InceptionV3, and MobileNet, for different datasets of UCM, SIRI-WHU, and RSSCN7 that have distinct parameters. In accuracy performance, the MobileNet outperformed the competition on the UCM and SIRI-WHU datasets, while EfficientNetB7 performed better on the RSSCN7 dataset. We also found that the dataset had an effect on the model's efficiency, with the UCM dataset outperforming the SIRI-WHU and RSSCN7 datasets across the board in terms of most measurement measures. The findings of this study indicated that it could provide significant benefits to remote sensing communities and decision-makers. The need for a powerful processing unit and the limited time frames caused by COVID-19 were the major challenges and limitations of this research. Based on these challenges, we have come up with some recommendations for the future, such as using a more powerful processor to improve the performance of DL models and applying DL hyperparameters to the domain area.