Innovative Use of High-Resolution Training Data and ResNet for Enhanced Land-Use Classification with Sentinel-2 Imagery

The classification of land cover and use (LULC) is essential for territorial government, and environmental planning. Satellite images obtained in the Copernicus European program are crucial in various fields of spatial management, particularly in the conservation of cultural and natural resources. This is largely due to their wide dissemination, allowing multiple institutions and organizations to utilize them. However, their low resolution can hinder the precision of supervised AI (Artificial Intelligence) classifiers, vital for continuous land-use monitoring, especially during the training phase. This phase can be costly owing to the need for advanced technology and extensive training datasets. The research focuses on the development of an AI classifier that leverages high-resolution training data and the powerful ResNet architecture with the Benchmark for Remote Sensing Image Classification (RSI-CB128). ResNet, known for its deep residual learning capabilities, significantly enhances the classifier's ability to learn complex patterns and features from high-resolution images. A testing dataset derived from Sentinel-2 raster images is utilised to analyse the performance of the neural network (NN). Our objectives are the evaluation and full confirmation of the effectiveness of an AI classifier applied to Sentinel-2 images produced with high-definition sensors. The findings indicate substantial improvements over current classification methods, such as U-Net and OBIA, underscoring the transformative potential of ResNet in advancing land-use classification accuracy.