Scholarlink

With the revolutionary development of machine learning and the significant growth of data volume, computer-aided change detection plays a critical role in many fields, such as Human Assistance and Disaster Response, critical region monitoring, environmental protection, agriculture, and forestry. Due to the complexity of change patterns and limited high-quality data, the existing systems and technologies make change detection out of reach for practical application, as recent reported performances testify in, for example, the state-of-the-art MapFormer approach. Change detection is a Dynamic Data Driven Applications System (DDDAS) paradigm example, which leverages available prior information from known models (e.g., terrain maps) and simulated models (e.g., buildings), while simultaneously considering the corresponding observed imagery from instrumentation. To develop the DDDAS damage assessment, an image-based pre- and post-conditional probability learning (IP2CL) algorithm design was introduced. It shows significant improvements over existing performances of damage level classification by using pixel-wise semantic and patch-based segmentation based contrastive learning. IP2CL encodes an image pair and known foreground mask into one image where the change domains are differentiated with the prior geospatial information resulting in accurate highlighting. This work demonstrates these improvements and the utility of IP2CL embedding in change detection tasks on the High Resolution Semantic Change Detection (HRSCD) dataset. Empirical results from our tests demonstrate a strong representational capability of IP2CL in its ability to encode the changes between two corresponding images into a single image. This transformation allows for the use of effective deep learning techniques, such as UNet and contrastive learning, bolstering high fidelity model accuracy, especially expressed when the size of the dataset is small. The higher performance and demands for a smaller training dataset make this methodology a valuable tool for practical applications.

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

DDDAS Probability Learning for Natural Disaster Change Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Weicong Feng,
  • Adarsh Agrawal,
  • Haibin Ling,
  • Erik Blasch,
  • Erika Adiles-Cruz,
  • Paul T. Schrader,
  • Jie Wei

摘要

With the revolutionary development of machine learning and the significant growth of data volume, computer-aided change detection plays a critical role in many fields, such as Human Assistance and Disaster Response, critical region monitoring, environmental protection, agriculture, and forestry. Due to the complexity of change patterns and limited high-quality data, the existing systems and technologies make change detection out of reach for practical application, as recent reported performances testify in, for example, the state-of-the-art MapFormer approach. Change detection is a Dynamic Data Driven Applications System (DDDAS) paradigm example, which leverages available prior information from known models (e.g., terrain maps) and simulated models (e.g., buildings), while simultaneously considering the corresponding observed imagery from instrumentation. To develop the DDDAS damage assessment, an image-based pre- and post-conditional probability learning (IP2CL) algorithm design was introduced. It shows significant improvements over existing performances of damage level classification by using pixel-wise semantic and patch-based segmentation based contrastive learning. IP2CL encodes an image pair and known foreground mask into one image where the change domains are differentiated with the prior geospatial information resulting in accurate highlighting. This work demonstrates these improvements and the utility of IP2CL embedding in change detection tasks on the High Resolution Semantic Change Detection (HRSCD) dataset. Empirical results from our tests demonstrate a strong representational capability of IP2CL in its ability to encode the changes between two corresponding images into a single image. This transformation allows for the use of effective deep learning techniques, such as UNet and contrastive learning, bolstering high fidelity model accuracy, especially expressed when the size of the dataset is small. The higher performance and demands for a smaller training dataset make this methodology a valuable tool for practical applications.