Landscape Classification Method Using Improved U-Net Model in Remote Sensing Image Ecological Environment Monitoring...

Area of Science:

• Remote Sensing
• Computer Vision
• Ecological Informatics

Background:

• Traditional remote sensing image classification methods suffer from low accuracy and are time-consuming.
• Accurate classification of ecological garden landscapes is crucial for dynamic monitoring and development assessment.

Purpose of the Study:

• To propose an improved U-Net model for accurate and efficient remote sensing image classification of ecological garden landscapes.
• To address challenges like model overfitting and incomplete detection of small targets in existing methods.

Main Methods:

• Collected remote sensing images using an s185 multirotor unmanned aerial vehicle (UAV) system.
• Preprocessed data using min-max standardization and data enhancement.
• Developed an Att-Unet network model incorporating asymmetric convolution blocks and an attention mechanism.
• Applied fully connected conditional random fields for postclassification refinement.

Main Results:

• Achieved a recall of 0.854, precision of 0.801, F1 value of 0.836, and accuracy of 0.982 for ecological garden landscape classification.
• Demonstrated a classification test time of 8.9 seconds.
• Outperformed other comparison methods in overall performance.

Conclusions:

• The proposed improved U-Net model (Att-Unet) offers superior performance for remote sensing image classification of ecological garden landscapes.
• The method provides effective theoretical support for the dynamic monitoring of ecological garden development.
• The integration of attention mechanisms and conditional random fields enhances classification accuracy and efficiency.