Land use classification using multi-year Sentinel-2 images with deep learning ensemble network

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence

Background:

• Accurate land use classification is crucial for urban planning, environmental monitoring, and agriculture.
• Sentinel-2 satellite imagery offers valuable spatial and spectral data for land cover analysis.
• Existing methods may lack robustness in multi-year classification tasks.

Purpose of the Study:

• To develop and evaluate a novel deep learning ensemble network, IRUNet, for multi-year land use classification.
• To enhance classification accuracy and robustness using Sentinel-2 imagery.
• To provide a generalizable framework for land use mapping.

Main Methods:

• Integration of InceptionResNetV2 with a UNet framework to create IRUNet.
• Application of multi-scale feature fusion for improved data representation.
• Utilization of Test-Time Augmentation (TTA) to increase prediction robustness.
• Classification of multi-year Sentinel-2 imagery for the Katpadi region (2017-2024).

Main Results:

• IRUNet achieved a high accuracy of 98.21% and a Dice Similarity Coefficient (DSC) of 88.96%.
• The model demonstrated superior performance compared to UNet, ResUNet, and Attention-UNet.
• Precision (94.71%) and recall (89.19%) metrics further validate the model's effectiveness.
• The study reported additional metrics including F1-score and Kappa coefficient.

Conclusions:

• IRUNet presents a high-performance and generalizable deep learning framework for multi-year land use classification.
• The proposed method effectively leverages Sentinel-2 data for detailed land cover mapping.
• The findings support the application of IRUNet in urban planning and environmental management.