|
| 基于CNN-LSTM耦合模型的禹城市土壤有机质预测制图 |
| Soil organic matter prediction and mapping in Yucheng City based on CNN-LSTM coupled model |
| 投稿时间：2025-03-05 |
| DOI：10.13254/j.jare.2025.0161 |
| 中文关键词: 土壤有机质，数字土壤制图，CNN-LSTM，环境协变量，MODIS物候数据，时空特征融合，禹城市 |
| 英文关键词: soil organic matter, digital soil mapping, CNN-LSTM, environmental covariate, MODIS phenological data, spatiotemporal feature fusion, Yucheng City |
| 基金项目:2021年山东省高等学校“青创人才引育计划”项目；山东省高等学校青创科技支持计划项目（2024KJH092） |
|
| 摘要点击次数: 584 |
| 全文下载次数: 68 |
| 中文摘要: |
| 当前数字土壤制图（DSM）研究在环境变量选择中对物候信息系统考量不足，现有模型时空多维特征融合能力有限。为探索融合MODIS长时间序列物候数据的新方法，提升耕地土壤有机质（SOM）的预测精度，本研究以山东省禹城市为研究区，整合环境协变量与2011—2020年MODIS物候数据，通过卷积神经网络（CNN）提取空间特征，结合长短期记忆网络（LSTM）捕捉时序动态规律，创新性构建CNN-LSTM时空特征耦合模型以实现耕地SOM的高精度预测。结果表明：融合物候特征使CNN-LSTM模型预测精度显著提升（R2=0.523），较未使用物候数据的CNN模型（R2=0.492）和随机森林模型（R2=0.467）精度分别提高6.3%和12.0%；SOM空间分布呈中南部高、北部及东南部低格局，与地形特征和耕作模式高度吻合；风速、水汽压和地形位置指数与SOM显著相关，物候变量中生长季中期的植被指数对预测贡献度最高。通过将MODIS长时间序列物候数据与CNN-LSTM时空融合模型相结合，本研究不仅揭示了物候动态对SOM预测的增强作用，同时有效突破了传统方法在土壤属性时空特征融合中的技术瓶颈，所生成的30 m分辨率SOM数字土壤制图可为区域耕地质量提升和土壤固碳潜力评估提供科学数据支持。 |
| 英文摘要: |
| To address the insufficient systematic consideration of phenological information in environmental variable selection for digital soil mapping（DSM）and the limited spatiotemporal multidimensional feature fusion capabilities of existing models, this study aims to explore a novel method by integrating moderate resolution imaging spectroradiometer（MODIS）long-term time-series phenological data to improve the prediction accuracy of soil organic matter（SOM） in cropland areas. Focusing on Yucheng City, Shandong Province, environmental covariates and MODIS phenological data from 2011 to 2020 were integrated, with spatial features extracted using a convolutional neural network（CNN） and temporal dynamic patterns captured with a long short-term memory（LSTM） network. An innovatively designed CNN-LSTM spatiotemporal feature coupling model was developed to achieve high-accuracy SOM prediction. Results indicated that the integration of phenological features significantly improved the prediction accuracy of the CNN-LSTM model（R2=0.523）, with 6.3% and 12.0% enhancements compared with CNN（R2=0.492）and random forest model（R2=0.467）without phenological data, respectively. The spatial distribution of SOM exhibited a pattern of higher values in central and south regions and lower values in northern and southeastern areas, strongly aligned with local topography and farming practices. Environmental factors such as wind speed, vapor pressure, and topographic position index showed significant correlations with SOM, while vegetation indices during mid-growth seasons contributed most to SOM prediction among phenological variables. By combining MODIS long-term phenological data with the CNN-LSTM spatiotemporal fusion model, this study not only demonstrates the enhanced role of phenological dynamics in SOM prediction but also overcomes the technical limitations of traditional methods in spatiotemporal feature fusion for soil attributes. The generated 30-meter-resolution SOM digital soil mapping product provides robust scientific data support for regional cropland quality improvement and soil carbon sequestration potential assessment. |
| HTML
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
