|
| 基于多时相Sentinel-2影像与模型集成的土壤质地遥感反演 |
| Mapping soil texture using multi-temporal Sentinel-2 imagery and ensemble modeling |
| 投稿时间：2026-02-27 |
| DOI：10.13254/j.jare.2026.0199 |
| 中文关键词: 土壤质地，多时相遥感数据，模型集成，黄河三角洲 |
| 英文关键词: soil texture, multi-temporal remote sensing data, ensemble modeling, Yellow River Delta |
| 基金项目: |
|
| 摘要点击次数: 367 |
| 全文下载次数: 47 |
| 中文摘要: |
| 为明确多时相遥感数据与模型集成在土壤质地反演中的作用，进一步提高黄河三角洲地区土壤质地反演精度，本研究基于统一环境协变量，融合不同时相哨兵二号（Sentinel-2）多光谱指数，构建了7组特征数据集。基于不同数据集，运用随机森林（RF）、极端梯度提升（XGBoost）、支持向量回归（SVR）、岭回归（Ridge）和轻量级梯度提升机（LightGBM）五种模型分别对砂粒、粉粒和黏粒含量进行了建模对比，并深入探讨了单时相跨模型集成与多时相数据集成两种策略对土壤质地反演的改进作用。结果表明，遥感影像时相选择显著影响土壤质地反演精度，即使在相同作物物候阶段内，不同成像日期的影像反演效果仍存在差异。单模型对比显示，树模型（RF、XGBoost、LightGBM）在砂粒和粉粒预测中明显优于Ridge和SVR，而黏粒预测中SVR与树模型精度相当，Ridge性能也显著提升。集成方面，堆叠集成（Stacking）优于简单平均集成。单时相多模型堆叠集成相较于最优单一模型砂粒预测R2稳定提升0.02~0.03，粉粒和黏粒未表现出明显优势；多时相堆叠集成效果较优，砂粒、粉粒、黏粒的R2相比最优单模型分别提升0.12、0.08、0.05。研究表明，基于多时相遥感信息的堆叠集成方法能有效提升土壤质地反演精度，在黄河三角洲这类成土环境复杂区域的高分辨率土壤数字制图中表现出良好适用性，可为类似区域土壤属性遥感反演提供方法参考。 |
| 英文摘要: |
| To clarify the role of multi-temporal remote sensing data and ensemble models in soil texture prediction and to further improve the prediction accuracy in the Yellow River Delta region, this study constructed seven feature datasets by integrating unified environmental covariates with Sentinel-2 multispectral indices from different acquisition dates. Using these datasets, the research applied five models, including Random Forest（RF）, Extreme Gradient Boosting（XGBoost）, Support Vector Regression（SVR）, Ridge Regression（Ridge）, and LightGBM, to predict sand, silt, and clay contents and compared their performance. In addition, the research investigated the effects of two ensemble strategies that integrate different models within a single temporal phase and combine information across multiple temporal datasets on soil texture prediction accuracy. Results showed that remote sensing image timing significantly affected the accuracy of soil texture inversion. Even within the same crop phenological stage, images acquired on different dates still exhibited differences in prediction performance. In single-model comparison, tree-based models（RF, XGBoost, LightGBM）were clearly superior to Ridge and SVR for sand and silt content. For clay content, SVR achieved comparable accuracy to the tree-based models, and Ridge performance also improved substantially. In terms of ensemble strategies, stacked ensemble generally outperformed simple averaging ensemble. For single-temporal datasets, multi-model stacked ensemble consistently improved the validation R2 for sand by 0.02-0.03 relative to the best-performing single model. The multi-temporal stacked ensemble achieved superior performance, yielding R2 improvements of 0.12, 0.08, and 0.05 for sand, silt, and clay, respectively, compared with the best single model. These results indicate that stacked ensemble methods based on multi-temporal remote sensing information can improve the accuracy of soil texture prediction. They demonstrate strong applicability in high-resolution digital soil mapping in complex pedogenic environments such as the Yellow River Delta and may provide a useful methodological reference for soil property retrieval in similar regions. |
| HTML
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
