Comparison of CNN Algorithms on Hyperspectral Image Classification in Agricultural Lands

Area of Science:

• Agricultural remote sensing
• Machine learning for image analysis
• Hyperspectral imaging applications

Background:

• Hyperspectral images (HSIs) offer rich spectral information for agricultural land classification.
• Convolutional Neural Networks (CNNs) have shown promise in analyzing complex HSI data.
• Developing efficient CNN models is crucial for accurate agricultural monitoring.

Purpose of the Study:

• To evaluate and compare various CNN architectures for hyperspectral agricultural land classification.
• To identify the most effective CNN approach for HSI data analysis in agriculture.
• To achieve high classification accuracy for crop and vegetation types.

Main Methods:

• Development and comparison of several CNN versions: 1D-CNN (pixelwise spectral, selected bands, spectral-spatial features) and 2D-CNN (principal components).
• Utilized HSI datasets from crop agriculture (Salinas Valley) and mixed vegetation (Indian Pines).
• Augmented input vectors for 1D-CNN to incorporate both spectral and spatial features.

Main Results:

• The 1D-CNN with augmented input vectors, integrating spectral-spatial features, achieved the highest accuracies.
• Classification accuracies reached 99.8% for the Salinas Valley dataset and 98.1% for the Indian Pines dataset.
• This approach outperformed other tested CNN versions in hyperspectral image classification.

Conclusions:

• The 1D-CNN model with augmented spectral-spatial features is highly effective for classifying agricultural hyperspectral images.
• Integrating both spectral and spatial information within CNNs significantly enhances classification performance.
• This study provides a robust method for accurate agricultural land cover mapping using HSIs.