Scholarlink

<p>Time-series clustering plays a critical role in identifying hidden patterns and similarities in complex datasets over time especially when labelled data are not available. This task becomes even more challenging when the data are high-dimensional, such as remote sensing images that are usually both multi-temporal and multi-spectral. There are various deep learning-based approaches, but they usually have complex training phase, and many convolution-based models require learnable kernels for feature extraction from time-series data. In this work, we have designed a clustering framework that uses static random kernels to transform time-series data of Sentinel-2 using MiniRocket (MINImally RandOm Convolutional KErnel Transform). The extracted features serve as input to clustering algorithm for agricultural land segmentation. We tested the performance of two clustering algorithms: K-Means and dfkmeans-<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(l^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>l</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> in terms of widely used clustering validation indices. In addition, we have also tested the clustering quality of the dataset based on features extracted from TS2Vec and tsfresh. Experimental results demonstrate that the proposed framework has great potential to be used for various downstream tasks and applications, which works on high-dimensional datasets with no labels and require timely outcomes for decision-making.</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

Clustering framework based on random convolutional kernels for segmentation of time-series satellite imagery

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Maleeha Najam,
  • Hasnat Khurshid,
  • Faisal Akram

摘要

Time-series clustering plays a critical role in identifying hidden patterns and similarities in complex datasets over time especially when labelled data are not available. This task becomes even more challenging when the data are high-dimensional, such as remote sensing images that are usually both multi-temporal and multi-spectral. There are various deep learning-based approaches, but they usually have complex training phase, and many convolution-based models require learnable kernels for feature extraction from time-series data. In this work, we have designed a clustering framework that uses static random kernels to transform time-series data of Sentinel-2 using MiniRocket (MINImally RandOm Convolutional KErnel Transform). The extracted features serve as input to clustering algorithm for agricultural land segmentation. We tested the performance of two clustering algorithms: K-Means and dfkmeans- \(l^2\) l 2 in terms of widely used clustering validation indices. In addition, we have also tested the clustering quality of the dataset based on features extracted from TS2Vec and tsfresh. Experimental results demonstrate that the proposed framework has great potential to be used for various downstream tasks and applications, which works on high-dimensional datasets with no labels and require timely outcomes for decision-making.