Research on Crop Classification Using U-Net Integrated with Multimodal Remote Sensing Temporal Features.

Area of Science:

• Agricultural remote sensing
• Geospatial analysis
• Machine learning for agriculture

Background:

• Accurate crop classification is crucial for food security and efficient agricultural management.
• Conventional methods often use single data sources, limiting temporal and spatial accuracy.
• Integrating optical and radar data offers complementary information for improved classification.

Purpose of the Study:

• To develop and evaluate a feature-level fusion method for crop classification using multimodal remote sensing data.
• To overcome limitations of single-sensor approaches by combining optical and SAR imagery.
• To enhance the accuracy and consistency of crop classification for corn and soybeans.

Main Methods:

• Feature extraction from Sentinel-2 optical and Sentinel-1 radar imagery.
• Identification of optimal feature combinations (NDVI+NDRE, VV+VH) using random forest.
• Feature-level fusion of 16 optical and 30 radar scenes into a 46-channel image.
• Crop classification using a U-Net deep neural network, compared to single-modal results.

Main Results:

• The multimodal fusion model achieved high classification accuracies: 95.83% (training), 91.99% (validation), and 90.81% (testing).
• Fusion model demonstrated superior performance over single-modal approaches in accuracy, boundary delineation, and consistency.
• Significant improvements were noted in F1-score, precision, and recall metrics.

Conclusions:

• Feature-level fusion of optical and radar remote sensing data provides a robust method for accurate crop classification.
• The proposed U-Net based fusion model effectively integrates multimodal data, outperforming traditional methods.
• This approach enhances agricultural monitoring capabilities, contributing to better resource management and food security.