Scholarlink

In this paper, we present the Shallow Parallel Convolutional Neural Network (SP-CNN), a novel method for classifying remote sensing images. Designed to work with sparse, point-based training data, SP-CNN leverages both spectral and spatial information by combining multiple shallow CNNs, each targeting different pixel neighborhoods. This parallel structure allows the model to capture essential contextual information, improving classification accuracy over traditional pixel-based methods like Support Vector Machines (SVM). In a crop mapping task, SP-CNN achieved a classification accuracy of 92%, surpassing SVM’s 90.6%. The flexibility of SP-CNN, particularly its capacity to adjust the number of parallel branches based on the relative size of the target object to the pixel size, demonstrates its applicability to various remote sensing challenges. While further optimization is needed to address computational complexity, SP-CNN represents a promising tool for enhancing the classification of medium- to high-resolution satellite imagery.

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

Shallow Convolutional Networks for Crop Mapping in the Context of Sparse Training Points

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Bassam Abdellatif

摘要

In this paper, we present the Shallow Parallel Convolutional Neural Network (SP-CNN), a novel method for classifying remote sensing images. Designed to work with sparse, point-based training data, SP-CNN leverages both spectral and spatial information by combining multiple shallow CNNs, each targeting different pixel neighborhoods. This parallel structure allows the model to capture essential contextual information, improving classification accuracy over traditional pixel-based methods like Support Vector Machines (SVM). In a crop mapping task, SP-CNN achieved a classification accuracy of 92%, surpassing SVM’s 90.6%. The flexibility of SP-CNN, particularly its capacity to adjust the number of parallel branches based on the relative size of the target object to the pixel size, demonstrates its applicability to various remote sensing challenges. While further optimization is needed to address computational complexity, SP-CNN represents a promising tool for enhancing the classification of medium- to high-resolution satellite imagery.