基于无人机多光谱信息与器官空间分布特征融合的油菜地上生物量估测

Remote Estimation of Rapeseed Above-ground Biomass by Fusion of UAV Multi-spectral Information and Spatial Distribution Feature of Organs

摘要:

地上生物量(Above-ground biomass, AGB)是作物长势评估和产量预测的重要指标之一。为构建基于无人机多光谱遥感的油菜AGB估测模型，本研究利用无人机搭载6波段多光谱传感器获取油菜苗期、蕾薹期、花期和角果期遥感影像，通过影像傅里叶变换提取频谱特征，研究不同生育期频谱特征和地上生物量之间的相关性，提取表征叶、花、角果等器官空间分布的频谱特征分量，并与光谱特征融合。然后，利用反向传播(Back propagation, BP)神经网络、随机森林(Random forest, RF)和梯度提升算法(eXtreme gradient boosting, XGBoost)3种机器学习算法构建不同生育期AGB估测模型，比较不同模型输入变量下AGB估测精度。结果表明：基于光谱和器官空间分布的融合特征构建AGB估测模型，可以提高AGB估测精度;3种机器学习算法中，BP更具稳定性，对融合特征有较好的AGB拟合效果;以6个波段反射率和频谱特征（叶期、角果期选取低频分量，花期选取高频分量，全生育期选取低频和高频加权重构分量）作为输入变量，结合BP算法能准确估测油菜不同生育期AGB，叶期测试集决定系数(R2)为0.87，均方根误差(RMSE)为178.44 g/m2;花期测试集R2为0.66，RMSE为357.76 g/m2;角果期测试集R2为0.87，RMSE为258.71 g/m2;全生育期测试集R2为0.79，RMSE为388.90 g/m2。研究结果可为精准、高效获取区域尺度油菜AGB提供技术支撑。

Abstract:

Above-ground biomass (AGB) is one of the most important indicators for crop growth assessment and yield prediction. To establish the rapeseed AGB estimation model based on UAV multispectral data, a UAV equipped with a 6-band multispectral sensor was utilized to acquire remote sensing images during the seedling, elongation, flowering, and pod stages of rapeseed. The Fourier transform was conducted to UAV multispectral images, and frequency-domain features were extracted based on Fourier spectrum. Then, the correlation between the frequency-domain features and AGB was analyzed at different growth stages, and the frequency components representing the spatial distribution of organs (leaves, flowers, and pods) were extracted and fused with spectral features. Three machine learning algorithms of back propagation (BP) neural network, random forest (RF), and eXtreme gradient boosting (XGBoost) were employed to establish AGB estimation models for different growth stages and the accuracy of AGB estimation with different input variables was compared. The results indicated that the fusion of spectral features and organ spatial distribution features can improve the accuracy of AGB estimation model. Among the three machine learning algorithms, BP demonstrated more stable and better AGB fitting performance with the fused features. Using the reflectance of the six bands and specific frequency-domain features (low-frequency components for the leaf and pod stages;high-frequency components for the flowering stage;weighted reconstruction of low and high frequency components for the entire growth stage) as input variables, BP achieved accurate estimation of rapeseed AGB across different growth stages. The test set results were as follows, coefficient of determination (R2) of leaf stage was 0.87 with root mean square error (RMSE) of 178.44 g/m2, R2 of flowering stage was 0.66 with RMSE of 357.76 g/m2, R2 of pod stage was 0.87 with RMSE of 258.71 g/m2 and R2 of entire growth period was 0.79 with RMSE of 388.90 g/m2. The results can provide technical support for the precise and efficient acquisition of rapeseed AGB at a regional scale.