Effective Sequential Classifier Training for SVM-Based Multitemporal Remote Sensing Image Classification.

Area of Science:

• Remote Sensing
• Machine Learning
• Image Classification

Background:

• Supervised classification algorithms like Support Vector Machines (SVM) struggle with limited training samples in remote sensing due to costly ground truthing.
• Multitemporal remote sensing images offer potential solutions through temporal correlation and spectral similarity.

Purpose of the Study:

• To propose a novel SVM-based Sequential Classifier Training (SCT-SVM) approach for enhanced multitemporal remote sensing image classification.
• To reduce the dependency on extensive ground-truthing by leveraging information from previous images in a sequence.

Main Methods:

• Developed a progressive SCT-SVM approach that utilizes classifiers from preceding images to train classifiers for incoming images.
• Implemented a two-step process: initial classifier prediction based on temporal trends, followed by fine-tuning with current training samples.
• Validated the method using Sentinel-2A multitemporal data over an agricultural region in Australia.

Main Results:

• The SCT-SVM approach significantly improved classification accuracy compared to state-of-the-art model transfer algorithms.
• In scenarios with insufficient training data, classification accuracy increased from 76.18% to 94.02% using SCT-SVM.
• Demonstrated the effectiveness of leveraging a priori information from previous images for subsequent classifications.

Conclusions:

• The proposed SCT-SVM method effectively addresses the challenge of limited training samples in multitemporal remote sensing image classification.
• Leveraging temporal information from prior images provides substantial benefits for classifying sequential remote sensing data.
• SCT-SVM offers a promising solution for accurate and efficient land cover classification using time-series satellite imagery.