Scholarlink

<p>This study introduces the crop and land cover land use (CLCLU) dataset, a 30 m resolution product providing annual maps of CLCLU across the transnational Middle Rio Grande (MRG) region, spanning both the U.S. and Mexico from 1994 to 2024. The model was trained using the Cropland Data Layer (CDL) on the US side. Dual-month (July and December) Landsat composites and a semantic segmentation model, MANet with ResNeXt-101 encoder, under four strategies were used to address sensor and temporal variability. This model architecture was chosen for its intrinsic ability to capture detailed spatial patterns and contextual dependencies through its attention-based design and ResNeXt-101 encoder, which demonstrated strong performance, particularly in generalizing across data-scarce regions in Mexico. The dataset achieved 97.10% overall accuracy and 78.85% mean Intersection over Union (mIoU), over validation process using a held-out CDL subset. Validation against NLCD and MCD12Q1-UMD confirmed high agreement. Data availability differences, minimal ground truth on the Mexican side, and cloud-related artifacts in early years led to some misclassification.</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

30 Years of simultaneous crop & land cover land use maps for Middle Rio Grande from 1994 to 2024

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月22日

查看全文
  • Saman Ebrahimi,
  • Mahdis Khorram,
  • Raquel Neri Barranco,
  • Rosario Sanchez,
  • Rocky Talchabhadel,
  • Santosh S. Palmate,
  • Marisol Dominguez-Tuda,
  • Elizabeth F. Racine,
  • Saurav Kumar

摘要

This study introduces the crop and land cover land use (CLCLU) dataset, a 30 m resolution product providing annual maps of CLCLU across the transnational Middle Rio Grande (MRG) region, spanning both the U.S. and Mexico from 1994 to 2024. The model was trained using the Cropland Data Layer (CDL) on the US side. Dual-month (July and December) Landsat composites and a semantic segmentation model, MANet with ResNeXt-101 encoder, under four strategies were used to address sensor and temporal variability. This model architecture was chosen for its intrinsic ability to capture detailed spatial patterns and contextual dependencies through its attention-based design and ResNeXt-101 encoder, which demonstrated strong performance, particularly in generalizing across data-scarce regions in Mexico. The dataset achieved 97.10% overall accuracy and 78.85% mean Intersection over Union (mIoU), over validation process using a held-out CDL subset. Validation against NLCD and MCD12Q1-UMD confirmed high agreement. Data availability differences, minimal ground truth on the Mexican side, and cloud-related artifacts in early years led to some misclassification.