Improved Agricultural Field Segmentation in Satellite Imagery Using TL-ResUNet Architecture.

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Growing global population and food security challenges necessitate efficient agricultural land monitoring.
• Precision agriculture relies on accurate land use classification and analysis for optimizing crop yields.
• Deep learning models offer superior performance in satellite image classification compared to traditional methods.

Purpose of the Study:

• To propose a novel deep learning model, Transfer Learning-based Residual UNet (TL-ResUNet), for semantic segmentation of land cover using satellite imagery.
• To enhance the accuracy and reliability of land cover classification in precision agriculture.

Main Methods:

• Developed a TL-ResUNet model integrating residual networks, transfer learning, and UNet architecture.
• Employed semantic segmentation for detailed land cover classification and mapping.
• Utilized public datasets, specifically DeepGlobe, for model evaluation.

Main Results:

• The TL-ResUNet model demonstrated superior performance compared to traditional models initialized with random weights and ImageNet pre-trained coefficients.
• Achieved a high Intersection over Union (IoU) score of 0.81 on the DeepGlobe validation dataset.
• Outperformed other models across multiple standard semantic segmentation metrics.

Conclusions:

• The proposed TL-ResUNet model is effective for accurate land cover classification and segmentation using satellite images.
• This advancement contributes to improved land use analysis and management in precision agriculture.
• The model's performance highlights the potential of deep learning for addressing global food security concerns.