Scholarlink

<p>The LUCAS land cover images data contains spatially explicit information on the land cover but due the limited number of images it presents a challenge in the training of classifiers. The primary objective of this research is to develop a robust classification framework that addresses the challenges of limited training data in ground-level land cover imagery through strategic combination of data augmentation, attention mechanisms, and ensemble learning strategies. This paper presents a comprehensive evaluation of this dataset by expanding the data with augmentation and then testing it on the state of the art convolutional neural network classifiers namely Densenet 121, EfficientNetB3, Xception and Resnet50. The study also tested Convolutional Block Attention Module embedded in these classifiers as well. Furthermore, this research introduces snapshot ensemble learning, which captures model checkpoints during training with cosine annealing learning rate schedules to create diverse ensembles from single training runs. This paper presents 20 classifiers in standard form, augmented with Convolutional Block Attention Module, snapshot ensembles, and traditional ensembles of these models as well. A comprehensive ablation study is also presented in this paper with in-depth analysis which shows that the proposed Snapshot EfficientNet-Vanilla achieves 95.74% accuracy and 95.73% F1-score, while the Smart Mixed Snapshot Ensemble achieves 95.75% accuracy and 95.67% F1-score, outperforming all individual models and traditional ensemble configurations. The analysis reveals that CBAM attention mechanisms exhibit architecture-dependent effects, improving ResNet50 performance while slightly degrading DenseNet121 and EfficientNetB3 in individual configurations. Snapshot ensembles demonstrate substantial improvements over individual models in 62.5% of configurations, with the combination of architectural diversity and snapshot-based temporal diversity providing complementary benefits. This study provides the first comprehensive analysis of multi-architecture ensembles with snapshot learning and attention mechanisms for land cover image classification.</p>

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

Snapshot ensemble learning with attention-enhanced CNNs for ground-level land cover classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月21日

查看全文
  • Simranjit Singh,
  • Sahil Sharma,
  • Vijay Kumar,
  • Amita Nandal

摘要

The LUCAS land cover images data contains spatially explicit information on the land cover but due the limited number of images it presents a challenge in the training of classifiers. The primary objective of this research is to develop a robust classification framework that addresses the challenges of limited training data in ground-level land cover imagery through strategic combination of data augmentation, attention mechanisms, and ensemble learning strategies. This paper presents a comprehensive evaluation of this dataset by expanding the data with augmentation and then testing it on the state of the art convolutional neural network classifiers namely Densenet 121, EfficientNetB3, Xception and Resnet50. The study also tested Convolutional Block Attention Module embedded in these classifiers as well. Furthermore, this research introduces snapshot ensemble learning, which captures model checkpoints during training with cosine annealing learning rate schedules to create diverse ensembles from single training runs. This paper presents 20 classifiers in standard form, augmented with Convolutional Block Attention Module, snapshot ensembles, and traditional ensembles of these models as well. A comprehensive ablation study is also presented in this paper with in-depth analysis which shows that the proposed Snapshot EfficientNet-Vanilla achieves 95.74% accuracy and 95.73% F1-score, while the Smart Mixed Snapshot Ensemble achieves 95.75% accuracy and 95.67% F1-score, outperforming all individual models and traditional ensemble configurations. The analysis reveals that CBAM attention mechanisms exhibit architecture-dependent effects, improving ResNet50 performance while slightly degrading DenseNet121 and EfficientNetB3 in individual configurations. Snapshot ensembles demonstrate substantial improvements over individual models in 62.5% of configurations, with the combination of architectural diversity and snapshot-based temporal diversity providing complementary benefits. This study provides the first comprehensive analysis of multi-architecture ensembles with snapshot learning and attention mechanisms for land cover image classification.