Crop classification with UAV multispectral remote sensing, employing an enhanced ResNet50 residual network

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• ResNet50 offers rapid training and high accuracy for crop feature extraction.
• Multispectral data's band correlation and redundancy can reduce classification accuracy.
• Standard ResNet50 requires improvement for multi-band data and high-precision, small-scale classification.

Purpose of the Study:

• To identify the optimal spectral band combination for classifying flue-cured tobacco and maize.
• To enhance the ResNet50 model for improved performance in crop classification tasks.
• To evaluate the impact of architectural modifications on classification accuracy.

Main Methods:

• Modified the ResNet50 model by incorporating Batch Normalization (BN), pyramid pooling, and hidden layers.
• Experimented with seven different combinations of these enhancements.
• Evaluated classification performance using metrics like accuracy, precision, recall, kappa coefficient, and F1 score.

Main Results:

• The RGB+NIR+Edge band combination achieved the highest classification accuracy (94.48%) for flue-cured tobacco and maize.
• Solely introducing BN layers provided the most significant improvement among the tested strategies.
• The enhanced model demonstrated high precision (94.66%), recall (94.48%), and F1 score (94.49%).

Conclusions:

• The RGB+NIR+Edge band combination is optimal for classifying flue-cured tobacco and maize using enhanced ResNet50.
• Batch Normalization layers are crucial for improving ResNet50's performance in multispectral crop classification.
• The study provides a robust methodology for high-precision crop classification with deep learning models.