Scholarlink

<p>Prior applications of deep-learning object detection in archaeology are often critiqued for their limited capacity to generate anthropologically meaningful insights into human societies. We address this by integrating comparative anthropology and fieldwork with a YOLO-based detector to trace shifting irrigation management in Turpan, Xinjiang, at a scale unattainable through manual inspection. Trained on tiled KH-9 HEXAGON and Planet SkySat imagery, our model identifies upcast mounds of underground irrigation aqueducts (<i>kariz</i>) with a precision of 87.7%. Across four time slices between 1974 and 2019, the detector identified over 160,000 features within <i>ca</i>. 500 km<sup>2</sup>. Beyond familiar patterns of mass destruction and urban or agricultural encroachment from the 1980s onward, diachronic analyses of bounding-box shape, size, and distribution reveal an increasingly fragmented landscape, with clustered features becoming more actively maintained in recent years. We argue that this transformation reflects path dependence and inter-community collaboration in Turpanian oases. Our analyses illuminate how communities negotiate and transform technology amid regime shifts and alternative technological choices. Methodologically, our study explores pathways through which deep-learning models can be productively integrated into archaeological hermeneutics.</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

Pixelated Transformations: Tracing Changing Irrigation Management Through Deep-Learning-Based Object Detection on HEXAGON and SkySat Satellite Images

  • 期刊论文

  • 发表日期：

    2026年4月23日

查看全文
  • Casey Haoran Shi,
  • Cayleigh Elizabeth Haag

摘要

Prior applications of deep-learning object detection in archaeology are often critiqued for their limited capacity to generate anthropologically meaningful insights into human societies. We address this by integrating comparative anthropology and fieldwork with a YOLO-based detector to trace shifting irrigation management in Turpan, Xinjiang, at a scale unattainable through manual inspection. Trained on tiled KH-9 HEXAGON and Planet SkySat imagery, our model identifies upcast mounds of underground irrigation aqueducts (kariz) with a precision of 87.7%. Across four time slices between 1974 and 2019, the detector identified over 160,000 features within ca. 500 km2. Beyond familiar patterns of mass destruction and urban or agricultural encroachment from the 1980s onward, diachronic analyses of bounding-box shape, size, and distribution reveal an increasingly fragmented landscape, with clustered features becoming more actively maintained in recent years. We argue that this transformation reflects path dependence and inter-community collaboration in Turpanian oases. Our analyses illuminate how communities negotiate and transform technology amid regime shifts and alternative technological choices. Methodologically, our study explores pathways through which deep-learning models can be productively integrated into archaeological hermeneutics.