基于无人机多光谱植被指数和纹理指数的马铃薯植株氮含量估算方法

Estimation of Potato Plant Nitrogen Content Based on UAV Multispectral Vegetation Indices and Texture Indices

摘要:

氮素是作物生长的重要元素之一,其精准监测对优化施肥和提高产量至关重要。本研究基于无人机多光谱及其变化特征开展马铃薯植株氮含量估算研究。获取了马铃薯块茎形成期、块茎增长期和淀粉积累期的多光谱影像,利用灰度共生矩阵提取多光谱影像上马铃薯的纹理特征,利用最优植被指数形式构建纹理植被指数,建立了纹理植被指数与植株氮含量的相关关系,以植被指数、纹理植被指数及其融合特征作为输入参数,使用人工神经网络、多元线性回归和偏最小二乘回归实现了不同生育期的马铃薯植株氮含量模型构建,并进行了验证。结果表明:在3个生育期达到极显著相关(P<0.01)的植被指数与植株氮含量的相关性在0.374~0.848之间,纹理植被指数与植株氮含量的相关性在0.625~0.855之间;以植被指数融合纹理植被指数为输入参数在3 个生育期构建的马铃薯植株氮含量模型效果优于以植被指数及纹理植被指数为输入参数构建的植株氮含量模型,其建模的最大R2和最小RMSE 分别为0.859和0.228% ,可以显著提高模型的精度和稳定性。相对于多元线性回归和偏最小二乘算法,人工神经网络估算马铃薯植株氮含量的精度最高。以植被指数为输入参数,利用人工神经网络平均建模的R2和RMSE为0.806和0.231% ,以纹理植被指数为输入参数,利用人工神经网络平均建模的R2和RMSE为0.681和0.300%,以融合植被指数和纹理植被指数为输入参数,利用人工神经网络平均建模的R2和RMSE为0.836和0.221%。该研究可为马铃薯氮含量快速反演及精准施肥提供技术支撑。

Abstract:

Nitrogen is a critical nutrient element for crop growth, and its accurate monitoring is essential for optimizing fertilization and improving yield. Aiming to develop an approach for estimating potato plant nitrogen content (PNC) by using unmanned aerial vehicle (UAV)-based multispectral imagery and its change characteristics, multispectral images were acquired during three key growth stages: tuber formation, tuber growth, and starch accumulation stage. The gray-level co-occurrence matrix (GLCM) method was applied to extract texture features from the multispectral images. Optimal vegetation indices were selected to construct texture indices based on vegetation indices, and the correlation between the vegetation indices, texture vegetation indices and plant nitrogen content was established. Three types of input features, i. e., the vegetation indices, texture vegetation indices and a fusion of vegetation indices and texture vegetation indices were used to build and validate estimation models for PNC at each growth stage. Three regression algorithms were compared: artificial neural network, multiple linear regression and partial least squares regression. Model performance was evaluated using the coefficient of determination (R2) and root mean square error (RMSE). The results showed that during the three growth stages, the correlation between vegetation indices and plant nitrogen content reached a highly significant level (P<0.01) ranged from 0.374 to 0.848, while the correlation between texture vegetation indices and plant nitrogen content ranged from 0.625 to 0.855, indicating that texture information improved the correlation. The potato plant nitrogen content model constructed by using fusion of vegetation index and texture vegetation index as input features outperformed the model built by using vegetation index and texture vegetation index alone. The highest R2 and the smallest RMSE values for calibration and validation across the three stages were 0.859 and 0.228%, respectively, demonstrating that the fusion approach significantly enhanced model accuracy and stability. Compared with multiple linear regression and partial least squares algorithms, artificial neural networks had the highest accuracy in estimating nitrogen content in potato plants. When using vegetation index, texture vegetation index, and fusion vegetation index and texture vegetation index as input parameters, the average calibration R2 and RMSE obtained by ANN were 0.806 and 0.231%, 0.681 and 0.300%, and 0.836 and 0.221%, respectively, in three critical stages. The proposed method, which integrated multispectral vegetation indices with texture features and employed ANN modeling, provided a reliable and efficient approach for rapid inversion of potato nitrogen content. This technique can support precision fertilization management in potato production.