Scholarlink

Context <p>High-resolution land-cover maps are essential for ecological monitoring and landscape-level analysis, yet long-term high-resolution time-series products remain scarce. The National Agricultural Imagery Program (NAIP) provides highly valuable aerial imagery, but strong temporal heterogeneity across years, primarily caused by shifts in sensor characteristics, limits the ability to generate consistent multi-year land cover maps.</p> Objectives <p>This study aims to (1) develop an algorithm capable of producing spatially detailed and temporally coherent 1-m land-cover maps using the state-of-the-art GeoAI/Machine Learning (ML) tools based on time-series NAIP imagery, and (2) address cross-year sensor characteristic shifts without relying on historical training samples.</p> Methods <p>We designed a dual-track adaptive workflow that applies different strategies to NAIP imagery with different qualities. NAIP imagery collected during 2009–2017 has a higher quality than that collected during 2004–2008. Images from the high-quality years were classified using a foundation model pretrained with U-Net/ResNet-34 and refined with a Segment Anything Model (SAM) for accurate boundary delineation. Images collected in the earlier years were reconstructed using an NLCD-based spatiotemporal bridging and label back-casting pipeline. Accuracy was evaluated across six U.S. counties in North Carolina and Pennsylvania.</p> Results <p>The algorithm produced stable results across years, yielding overall accuracies of 0.887 (2014), 0.886 (2017), and 0.733 (2004) with Kappa statistics at 0.860, 0.831, and 0.764, respectively. Structure, Water, Wetland, and Cropland exhibited consistently high F1-scores, while performance in low-quality imagery remained coherent despite substantial sensor differences. These findings demonstrate that the algorithm maintains both spatial fidelity and temporal consistency across heterogeneous historical imagery.</p> Conclusions <p>This study shows using GeoAI/ML tools along with multiple sources of data can effectively produce consistent high-resolution multi-decade land-cover maps based on NAIP imagery. The approach developed in this study provides a scalable solution for generating high-resolution time series land-cover maps across the conterminous USA where NAIP imagery is available, supporting long-term land-change analyses and landscape-level planning.</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

Using GeoAI and machine learning tools for consistent high-resolution land cover mapping based on time-series NAIP imagery

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月25日

查看全文
  • Jie Liu,
  • Xusheng Tang,
  • Chao Wang,
  • Zhengxiao Yan,
  • Yuchen Dai,
  • Qi Zhang,
  • Conghe Song

摘要

Context

High-resolution land-cover maps are essential for ecological monitoring and landscape-level analysis, yet long-term high-resolution time-series products remain scarce. The National Agricultural Imagery Program (NAIP) provides highly valuable aerial imagery, but strong temporal heterogeneity across years, primarily caused by shifts in sensor characteristics, limits the ability to generate consistent multi-year land cover maps.

Objectives

This study aims to (1) develop an algorithm capable of producing spatially detailed and temporally coherent 1-m land-cover maps using the state-of-the-art GeoAI/Machine Learning (ML) tools based on time-series NAIP imagery, and (2) address cross-year sensor characteristic shifts without relying on historical training samples.

Methods

We designed a dual-track adaptive workflow that applies different strategies to NAIP imagery with different qualities. NAIP imagery collected during 2009–2017 has a higher quality than that collected during 2004–2008. Images from the high-quality years were classified using a foundation model pretrained with U-Net/ResNet-34 and refined with a Segment Anything Model (SAM) for accurate boundary delineation. Images collected in the earlier years were reconstructed using an NLCD-based spatiotemporal bridging and label back-casting pipeline. Accuracy was evaluated across six U.S. counties in North Carolina and Pennsylvania.

Results

The algorithm produced stable results across years, yielding overall accuracies of 0.887 (2014), 0.886 (2017), and 0.733 (2004) with Kappa statistics at 0.860, 0.831, and 0.764, respectively. Structure, Water, Wetland, and Cropland exhibited consistently high F1-scores, while performance in low-quality imagery remained coherent despite substantial sensor differences. These findings demonstrate that the algorithm maintains both spatial fidelity and temporal consistency across heterogeneous historical imagery.

Conclusions

This study shows using GeoAI/ML tools along with multiple sources of data can effectively produce consistent high-resolution multi-decade land-cover maps based on NAIP imagery. The approach developed in this study provides a scalable solution for generating high-resolution time series land-cover maps across the conterminous USA where NAIP imagery is available, supporting long-term land-change analyses and landscape-level planning.