|
| 基于Sentinel-2影像与遗传算法优化堆叠模型的耕地土壤有机质遥感反演 |
| Mapping soil organic matter in cropland using Sentinel-2 imagery and a genetic algorithm-optimized stacking model |
| 投稿时间：2026-02-27 |
| DOI：10.13254/j.jare.2026.0200 |
| 中文关键词: 土壤有机质，Sentinel-2，光谱指数，相关性分析，遗传算法，堆叠模型 |
| 英文关键词: soil organic matter, Sentinel-2, spectral indice, correlation analysis, genetic algorithm, stacking model |
| 基金项目:国家重点研发计划项目（2023YFD1900100） |
|
| 摘要点击次数: 335 |
| 全文下载次数: 69 |
| 中文摘要: |
| 针对现有土壤有机质遥感反演中单一模型表达能力有限且堆叠模型优化不足的问题，本研究以吉林省伊通满族自治县耕地为研究对象，结合哨兵二号（Sentinel-2）多光谱影像和地形因子，比较不同特征方案下的单模型、常规堆叠模型和遗传算法优化的堆叠模型，旨在评估地形因子、特征筛选和基础学习器组合优化对土壤有机质反演精度的影响。首先，基于Sentinel-2影像提取各波段反射率，构建归一化差分指数（NDI）、广义差分指数（GDI）、比值指数（RI）和乘积指数（PI），分析其与有机质的相关性；其次，分别构建光谱特征、光谱+地形特征和递归特征消除（RFE）筛选特征三类输入方案，并采用9种算法进行单模型建模比较；最后，以9种模型作为候选基础学习器，构建常规堆叠模型和遗传算法优化堆叠模型，并比较其预测性能。结果表明，有机质与可见光至近红外波段反射率及构建的光谱指数均呈显著相关，其中B4、B8波段对有机质敏感性较高。与常规堆叠模型相比，遗传算法优化的堆叠模型（SVR+ERT+RF+ADB+GBT）可改善预测效果，测试集决定系数（R2）为0.748、均方根误差（RMSE）为0.252，相对分析误差（RPD）为2.009。空间制图结果显示，研究区有机质呈明显空间异质性，高值区集中于东部及西北部局部，低值区零星分布于边缘地带，整体呈连续过渡特征。研究表明，遗传算法优化堆叠模型能够有效融合多学习器优势，改善土壤有机质遥感反演效果，为黑土区土壤肥力定量评估与数字化监测提供技术参考。 |
| 英文摘要: |
| Accurate estimation of soil organic matter（SOM）using remote sensing remains challenging due to the limited predictive performance of individual models and the insufficiently optimized combination of ensemble models. This study constructed a stacking model integrating multispectral Sentinel-2 imagery and topographic features for SOM estimation in Yitong Manchu Autonomous County, Jilin Province. Spectral reflectance was derived from Sentinel-2 bands, and a series of spectral indices, including the normalized difference index （NDI ）, generalized difference index（GDI ）, ratio index（RI）, and product index（PI）, were constructed to enhance the spectral sensitivity to SOM. Three feature input schemes were established, namely spectral features, spectral + topographic features, and recursive feature elimination（RFE）- selected features. Individual models, conventional stacking models, and genetic algorithm-optimized stacking models under different feature schemes were constructed and compared. The results showed significant correlations between SOM and both visible- to-near-infrared reflectance and constructed spectral indices, with B4 and B8 bands exhibiting the strongest sensitivity. Compared with the conventional stacking model, the genetic algorithm- optimized stacking model（SVR+ERT+RF+ADB+GBT）further improved prediction accuracy（R2=0.748, RMSE=0.252, RPD=2.009）, indicating improved predictive performance. Spatial mapping revealed distinct heterogeneity in SOM distribution, with high-value areas mainly in the eastern and northwestern regions and a gradual transition across the county. These findings confirm that the genetic algorithm-optimized stacking framework effectively integrates multiple learning algorithms, improving SOM inversion accuracy and offering methodological support for quantitative soil fertility assessment and digital monitoring in black soil regions. |
| HTML
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
