Li Ming, Wang Weiyi, Li Haoran, Yang Zekun, Li Jianjun
School of Energy and Transportation Engineering, Inner Mongolia Agricultural University, Hohhot, China.
Front Plant Sci. 2025 Aug 13;16:1643945. doi: 10.3389/fpls.2025.1643945. eCollection 2025.
The rapid and accurate acquisition of vegetation information, particularly chlorophyll content, is essential for effective vegetation management and ensuring the safe operation of photovoltaic power plants. In this study, the vegetation within a photovoltaic power plant served as the research subject, and multispectral images were acquired using unmanned aerial vehicles, while chlorophyll measurements were obtained through ground-based sampling at multiple time points. From these images, twenty vegetation indices and thirty-two texture features were extracted. To reduce feature redundancy and enhance modeling efficiency, feature selection was performed using the minimum redundancy maximum relevance method and Pearson correlation analysis. The selected features were then used in three modeling strategies-vegetation index-based, texture feature-based, and fused index-texture-based-employing three conventional machine-learning regressors (partial least squares regression, random forest, support vector machine regression) and three deep-learning regressors (back propagation neural network, convolutional neural network, multilayer perceptron). Based on the optimal model, a spatiotemporal distribution map of chlorophyll content within the study area was generated. The results indicated that both vegetation indices and texture features exhibited significant correlations with chlorophyll content, with the strongest correlation observed between the green normalized difference vegetation index (GNDVI) and the NIR_Mean (Pearson coefficients of 0.82 and 0.65, respectively). Moreover, the fusion of vegetation indices and texture features effectively improved the accuracy of chlorophyll inversion models; among the six regression algorithms tested, the multilayer perceptron model achieved the highest performance (R² = 0.874, RMSE = 3.725, MAPE = 3.982%). This study provides a novel method for monitoring chlorophyll content in vegetation within photovoltaic power plant regions and offers informational support for refined regional vegetation management.
快速准确地获取植被信息，特别是叶绿素含量，对于有效的植被管理和确保光伏电站的安全运行至关重要。在本研究中，以光伏电站内的植被为研究对象，使用无人机获取多光谱图像，同时通过多个时间点的地面采样获得叶绿素测量值。从这些图像中提取了20种植被指数和32个纹理特征。为了减少特征冗余并提高建模效率，使用最小冗余最大相关性方法和Pearson相关分析进行特征选择。然后将所选特征用于三种建模策略——基于植被指数的、基于纹理特征的以及融合指数-纹理的——采用三种传统机器学习回归器（偏最小二乘回归、随机森林、支持向量机回归）和三种深度学习回归器（反向传播神经网络、卷积神经网络、多层感知器）。基于最优模型，生成了研究区域内叶绿素含量的时空分布图。结果表明，植被指数和纹理特征均与叶绿素含量呈现显著相关性，其中绿色归一化差异植被指数（GNDVI）与近红外均值（NIR_Mean）之间的相关性最强（Pearson系数分别为0.82和0.65）。此外，植被指数和纹理特征的融合有效地提高了叶绿素反演模型的精度；在所测试的六种回归算法中，多层感知器模型表现最佳（R² = 0.874，RMSE = 3.725，MAPE = 3.982%）。本研究为监测光伏电站区域内植被的叶绿素含量提供了一种新方法，并为精细化区域植被管理提供了信息支持。
