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

Ferdous Sohel, Yiqing Guo, Xiuping Jia, David PaullIEEE TIP 2018 · DOI

TL;DR

This paper introduces a novel Sequential Classifier Training for SVM (SCT-SVM) approach to tackle the challenge of limited training data in multitemporal remote sensing image classification. By leveraging knowledge from previously classified images, SCT-SVM significantly improves classification accuracy, especially when training samples are scarce.

What Problem Does This Paper Address?

Remote sensing technologies generate vast amounts of image data, but classifying this data accurately using supervised methods like Support Vector Machines (SVMs) is hampered by the laborious and expensive process of collecting ground truth data. This problem is exacerbated in agricultural areas where land cover changes rapidly due to crop phenology and human activities, quickly rendering existing ground truth obsolete. Traditional methods require substantial training data for each new image, which is often impractical. This paper aims to alleviate this by effectively utilizing the temporal correlation and spectral similarity between sequential remote sensing images to reduce the need for extensive new training data.

Methods

The core of the proposed approach is the Sequential Classifier Training for SVM (SCT-SVM). This method operates in two main stages:

  1. Classifier Prediction: For an incoming image, a rough initial classifier is predicted based on the temporal trend of a set of previously trained classifiers. This prediction leverages Principal Component Analysis (PCA) to capture the dominant temporal dynamics of classifier parameters. A polynomial regression is then used to model this temporal trend and predict the classifier for the current image.
  2. Classifier Fine-Tuning: The predicted classifier is then refined using a small set of training samples from the current image. This fine-tuning step employs a newly developed domain adaptation algorithm called Temporal-Adaptive SVM (TA-SVM). TA-SVM balances the information from the predicted classifier and the current training samples to achieve a more accurate final classifier. It aims to maximize the margin space while adhering to constraints that prevent the fine-tuned classifier from deviating too far from the predicted one.

The key innovation lies in transferring knowledge at the classifier level (i.e., classifier parameters) rather than the data level (i.e., raw pixels and labels), which reduces computational load and allows for classification even when raw data from previous images is inaccessible.

Datasets

The study utilizes two primary datasets:

  • Sentinel-2A Multitemporal Data: Collected over an agricultural area in Australia, this dataset consists of five cloud-free optical images from January to March 2017. It was used for the main experiments and classification of five land cover classes (crops, forests, water, soil, buildings).
  • Landsat-8 Multitemporal Data: Ten cloud-free Landsat-8 images from December 2016 to March 2017 were used for error analysis, specifically focusing on the rice and non-rice subclasses.

Key Insights and Ideas

  • Leveraging Temporal Dynamics: The temporal evolution of classifier parameters, not just spectral features, carries valuable information that can predict future classifier states.
  • Classifier-Level Transfer: Transferring knowledge at the classifier parameter level is more efficient and flexible than data-level transfer, especially when historical raw data is unavailable.
  • Domain Adaptation is Crucial: Direct application of previous classifiers is not feasible due to "dataset shifts" (changes in spectral characteristics). Domain adaptation techniques like TA-SVM are essential to bridge this gap.
  • Robustness to Errors: The SCT-SVM approach, particularly after fine-tuning, demonstrates robustness to errors in initial classifiers, becoming more reliable as more sequential images are processed.

Key Results

  • Significant Accuracy Improvement: When training data was insufficient (5 samples per class), the SCT-SVM approach improved overall classification accuracy from 76.18% to 94.02% compared to direct SVM application.
  • TA-SVM Superiority: The proposed TA-SVM fine-tuning algorithm outperformed existing methods like A-SVM and PMT-SVM, especially under limited training data conditions.
  • Optimal Fitting Functions: Second-order polynomial fitting was found to be most effective for predicting classifiers involving vegetative classes, while first-order fitting was better for non-vegetative classes.
  • Parameter Sensitivity: The regularization parameters F (weight for previous image information) and C (weight for current training samples) in TA-SVM significantly impact classification accuracy, with higher F values being crucial when training data is scarce.

Limitations and Future Work

  • Abrupt Changes: The current study does not explicitly handle abrupt changes in land cover (e.g., due to disasters like logging or forest fires), which might require defining new classes or stopping the sequential process. Future work could explore solutions for these cases, potentially incorporating active learning heuristics.
  • Other Classifiers: While the paper focuses on SVM, the authors suggest that similar concepts could be applied to other classifiers (e.g., Gaussian, Random Forests, Neural Networks), but specific algorithms would need to be developed to accommodate their unique characteristics.
  • Parameter Tuning: The optimal combination of regularization parameters (F and C) for TA-SVM requires careful tuning, often through cross-validation.

Summary updated

BibTeX
@article{sohel2018effective,
  author = {Ferdous Sohel and Yiqing Guo and Xiuping Jia and David Paull},
  title = {{Effective Sequential Classifier Training for SVM-Based Multitemporal Remote Sensing Image Classification}},
  journal = {IEEE TIP},
  year = {2018},
  doi = {10.1109/tip.2018.2808767},
}
Figure