Scholarlink

<p>The empirical retrieval of leaf area index (LAI) through multispectral remote sensing typically relies more on vegetation indices (VIs) than on partial least squares regression (PLSR). This is primarily because the limited number of bands in multispectral data may not fully leverage PLSR’s potential. However, multispectral remote sensing data provides more than just spectral reflectance, as new features can be extracted through various feature engineering techniques to satisfy the PLSR requirement. In this study, we developed PLSR models for LAI retrieval in a temperate deciduous forest with dramatic seasonal variations by integrating spectral bands, VIs, and texture features. We applied the least absolute shrinkage and selection operator (LASSO) and variance inflation factor (VIF) methods to select optimal features for the PLSR model. The results showed that the LASSO-based PLSR model achieved the best performance (R² = 0.640 ± 0.039). Red-edge and shortwave infrared reflectance remained highly sensitive to LAI, particularly in spring (R² = 0.636 ± 0.083) and autumn (R² = 0.837 ± 0.023). However, in summer, vegetation indices and texture features provided complementary structural information that compensated for the reduced sensitivity of reflectance alone. Despite strong seasonal variation, a generalized Spring–Autumn PLSR model was successfully established (R² = 0.726 ± 0.058). These findings demonstrate that integrating spectral reflectance with vegetation indices and textural features enhances the robustness and accuracy of multispectral LAI estimation.</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

Enhancing seasonal leaf area index estimation using PLSR with Sentinel-2 spectral and texture features: the importance of feature engineering and feature selection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月18日

查看全文
  • Xuanwen Wang,
  • Yi Gan,
  • Atsuhiro Iio,
  • Quan Wang

摘要

The empirical retrieval of leaf area index (LAI) through multispectral remote sensing typically relies more on vegetation indices (VIs) than on partial least squares regression (PLSR). This is primarily because the limited number of bands in multispectral data may not fully leverage PLSR’s potential. However, multispectral remote sensing data provides more than just spectral reflectance, as new features can be extracted through various feature engineering techniques to satisfy the PLSR requirement. In this study, we developed PLSR models for LAI retrieval in a temperate deciduous forest with dramatic seasonal variations by integrating spectral bands, VIs, and texture features. We applied the least absolute shrinkage and selection operator (LASSO) and variance inflation factor (VIF) methods to select optimal features for the PLSR model. The results showed that the LASSO-based PLSR model achieved the best performance (R² = 0.640 ± 0.039). Red-edge and shortwave infrared reflectance remained highly sensitive to LAI, particularly in spring (R² = 0.636 ± 0.083) and autumn (R² = 0.837 ± 0.023). However, in summer, vegetation indices and texture features provided complementary structural information that compensated for the reduced sensitivity of reflectance alone. Despite strong seasonal variation, a generalized Spring–Autumn PLSR model was successfully established (R² = 0.726 ± 0.058). These findings demonstrate that integrating spectral reflectance with vegetation indices and textural features enhances the robustness and accuracy of multispectral LAI estimation.