Scholarlink

Deploying deep learning models for multispectral satellite image classification in resource-constrained settings requires preserving predictive accuracy while reducing computation and storage. We benchmark two mainstream quantization strategies—Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) across convolutional (ResNet18, EfficientNet-B0) and transformer-based (Swin-Tiny, ConvMixer) architectures on EuroSAT and SIRI-WHU. Models are evaluated by accuracy, macro F1-score, model size, inference latency, and accuracy retention (relative to FP32 baselines). QAT consistently retains over 98% of full-precision accuracy with up to 5 \(\times \) compression, outperforming PTQ while incurring only modest fine-tuning overhead (5–10 epochs). Paired t-tests confirm the statistical significance of QAT gains over PTQ. While large-scale cloud platforms (e.g., Google Earth Engine) are widely used for processing multi-temporal remote sensing data, many land cover, land use applications require on-device inference due to connectivity, latency, or privacy constraints. In such scenarios, quantization provides a practical path to compact, fast, and accurate models. Our results offer actionable guidelines for selecting PTQ versus QAT under edge deployment constraints, particularly when spectral fidelity of multispectral inputs must be preserved.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Balancing Accuracy and Efficiency: Post-training Versus Quantization-Aware Training for Multispectral Remote Sensing Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Trinh Le Nhat,
  • Tan Thai Nhat,
  • Thao Nhien Hoang,
  • Luong Vuong Nguyen,
  • Cao Vu Bui

摘要

Deploying deep learning models for multispectral satellite image classification in resource-constrained settings requires preserving predictive accuracy while reducing computation and storage. We benchmark two mainstream quantization strategies—Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) across convolutional (ResNet18, EfficientNet-B0) and transformer-based (Swin-Tiny, ConvMixer) architectures on EuroSAT and SIRI-WHU. Models are evaluated by accuracy, macro F1-score, model size, inference latency, and accuracy retention (relative to FP32 baselines). QAT consistently retains over 98% of full-precision accuracy with up to 5 \(\times \) compression, outperforming PTQ while incurring only modest fine-tuning overhead (5–10 epochs). Paired t-tests confirm the statistical significance of QAT gains over PTQ. While large-scale cloud platforms (e.g., Google Earth Engine) are widely used for processing multi-temporal remote sensing data, many land cover, land use applications require on-device inference due to connectivity, latency, or privacy constraints. In such scenarios, quantization provides a practical path to compact, fast, and accurate models. Our results offer actionable guidelines for selecting PTQ versus QAT under edge deployment constraints, particularly when spectral fidelity of multispectral inputs must be preserved.