Scholarlink

<p>The dry–hot valley region is characterized by pronounced topographic relief and a complex land-use mosaic, resulting in strong nonlinearity and fine-scale spatial heterogeneity in soil organic matter (SOM). These characteristics increase the difficulty and uncertainty of remote sensing-based digital soil mapping. Although Yuanmou County has been widely investigated, limited attention has been paid to regional-scale SOM prediction in dry–hot valley ecosystems using long-term satellite-derived phenological dynamics and deep learning approaches. To address this gap, this study integrated 475 topsoil samples, topographic factors, environmental covariates, and multi-year satellite-derived phenology time-series variables to develop a regional-scale SOM prediction and assessment framework in Yuanmou County, Yunnan Province, China. The predictive performance of four models was compared, including a convolutional neural network (CNN), a CNN–random forest hybrid model (CNN-RF), a CNN–long short-term memory model (CNN-LSTM), and an attention-augmented CNN-LSTM model (CNN-LSTM-Att). The results showed that model architecture had a substantial influence on SOM prediction accuracy, with overall performance ranked as CNN-LSTM-Att &gt; CNN-LSTM &gt; CNN-RF &gt; CNN. Among the four models, CNN-LSTM-Att achieved the best performance on the independent test set, with R<sup>2</sup> = 0.61, RMSE = 2.49&#xa0;g&#xa0;kg⁻<sup>1</sup>, and MAE = 1.39&#xa0;g&#xa0;kg⁻<sup>1</sup>. The incorporation of temporal modeling and the attention mechanism improved the extraction of dynamic phenological signals associated with SOM formation, resulting in a more refined spatial representation. The CNN-LSTM-Att prediction map clearly identified low-SOM areas along the northern valley corridor and high-SOM zones in the forest-dominated southern and eastern regions, while showing stronger sensitivity to local patches and transitional ecotones. Overall, coupling long-term phenological dynamics with an attention mechanism improved both the predictive accuracy and spatial expressiveness of SOM mapping in complex terrain, providing a useful methodological reference for SOM assessment and precision land management in dry–hot valley regions.</p> Graphical Abstract <p></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

Regional-scale prediction and assessment of soil organic matter content in the Yuanmou Dry-Hot Valley, Southwest China, using satellite-derived phenological time series and deep learning

  • 期刊论文

  • 发表日期：

    2026年7月12日

查看全文
  • Dengdeng Ding,
  • Jida Yang,
  • Sihe Deng,
  • Hongye Zhu,
  • Zhengbo Ma,
  • Chengxiu Fu,
  • Junlang Deng,
  • Dingyun Zhao,
  • Zhen Long,
  • Xiaoyan Wang,
  • Peiwen Yang,
  • Qibin Chen,
  • Qing Zhang

摘要

The dry–hot valley region is characterized by pronounced topographic relief and a complex land-use mosaic, resulting in strong nonlinearity and fine-scale spatial heterogeneity in soil organic matter (SOM). These characteristics increase the difficulty and uncertainty of remote sensing-based digital soil mapping. Although Yuanmou County has been widely investigated, limited attention has been paid to regional-scale SOM prediction in dry–hot valley ecosystems using long-term satellite-derived phenological dynamics and deep learning approaches. To address this gap, this study integrated 475 topsoil samples, topographic factors, environmental covariates, and multi-year satellite-derived phenology time-series variables to develop a regional-scale SOM prediction and assessment framework in Yuanmou County, Yunnan Province, China. The predictive performance of four models was compared, including a convolutional neural network (CNN), a CNN–random forest hybrid model (CNN-RF), a CNN–long short-term memory model (CNN-LSTM), and an attention-augmented CNN-LSTM model (CNN-LSTM-Att). The results showed that model architecture had a substantial influence on SOM prediction accuracy, with overall performance ranked as CNN-LSTM-Att > CNN-LSTM > CNN-RF > CNN. Among the four models, CNN-LSTM-Att achieved the best performance on the independent test set, with R2 = 0.61, RMSE = 2.49 g kg⁻1, and MAE = 1.39 g kg⁻1. The incorporation of temporal modeling and the attention mechanism improved the extraction of dynamic phenological signals associated with SOM formation, resulting in a more refined spatial representation. The CNN-LSTM-Att prediction map clearly identified low-SOM areas along the northern valley corridor and high-SOM zones in the forest-dominated southern and eastern regions, while showing stronger sensitivity to local patches and transitional ecotones. Overall, coupling long-term phenological dynamics with an attention mechanism improved both the predictive accuracy and spatial expressiveness of SOM mapping in complex terrain, providing a useful methodological reference for SOM assessment and precision land management in dry–hot valley regions.

Graphical Abstract