Hyperspectral Remote Sensing Image Classification Based on Maximum Overlap Pooling Convolutional Neural Network.

Area of Science:

• Computer Science
• Remote Sensing
• Image Processing

Background:

• Traditional convolutional neural networks often use non-overlapping average pooling, which can lead to feature similarity and loss of detail in hyperspectral images.
• This loss of detail hinders accurate image classification, particularly for continuous spectrum hyperspectral data.

Purpose of the Study:

• To propose and evaluate a novel overlapping pooling method for convolutional neural networks to improve hyperspectral image classification.
• To address the limitations of average pooling in extracting distinct and detailed features from hyperspectral imagery.

Main Methods:

• An improved convolutional neural network architecture was developed, incorporating an overlapping pooling strategy.
• Maximum pooling was utilized with a smaller step size than the pooling kernel size to ensure overlap and coverage.
• The Indian Pines dataset, acquired by the airborne visible/infrared imaging spectrometer (AVIRIS) sensor, was used for experimental validation.

Main Results:

• The proposed overlapping pooling method effectively improved the extraction of image details in hyperspectral remote sensing data.
• Experimental results demonstrated a significant enhancement in classification accuracy when using the improved convolutional neural network.

Conclusions:

• The novel overlapping pooling method offers a superior approach for hyperspectral image classification compared to traditional non-overlapping pooling.
• This technique enhances feature extraction, leading to higher accuracy and better preservation of image details in remote sensing applications.