基于无人机多源遥感的玉米LAI垂直分布估算

为探究无人机多源遥感影像估算玉米叶面积指数(Leaf area index, LAI)垂直分布,在田间设置了密度和播期试验,在7个生育时期利用无人机采集了可见光、多光谱和热红外影像并同步获取玉米LAI垂直分布数据。同时,为合理制定无人机飞行任务,分析了不同飞行高度和不同太阳高度角下获取的无人机影像对估算玉米LAI的影响。基于无人机影像提取的与玉米LAI相关性较高的植被指数、纹理信息和冠层温度等特征,利用7种机器学习方法分别构建了玉米冠层不同高度LAI估算模型,从中选取鲁棒性强的2个模型用于分析在不同飞行高度和不同太阳高度角下估算LAI的差异。研究结果表明,MLPR和RFR模型对玉米LAI估算鲁棒性最强,全生育期下模型rRMSE为11.31%(MLPR)和11.42%(RFR)。玉米冠层LAI垂直分布估算误差,所有模型的平均rRMSE分别为9.1%(LAI-1)、14.19%(LAI-2)、18.62%(LAI-3)、23.29%(LAI-4)和26.7%(LAI-5)。对于玉米穗位叶及以下部位的LAI估算误差均在20%以下,得到了较好精度。同时,在不同飞行高度和太阳高度角试验中可以得出...

Vertical Distribution Estimation of Maize LAI Using UAV Multi-source Remote Sensing

Maize leaf area index (LAI) displays a significant vertical distribution gradient. However, there is currently a limited amount of research focused on directly estimating the vertical distribution patterns of maize LAI from images. Designing an appropriate unmanned aerial vehicle (UAV) detection scheme can contribute to improving the accuracy of maize LAI estimation. Thus different maize varieties, and density and disease were used, and sowing experiments were carried out in the field to collect data on the vertical distribution of maize LAI. UAVs equipped with RGB, multi-spectral (MS), and thermal infrared (TIR) cameras were used to capture visible, multi-spectral, and thermal infrared images. Seven sets of UAV image data were collected during the reproductive growth stage of maize. To validate the effects of different UAV flight altitudes and solar zenith angles on maize LAI estimation, two completely controlled experiments with different flight altitudes were conducted, resulting in a total of 10 sets of UAV image data. Additionally, UAV image data were collected at each hour from 08:00 to 18:00 on a single day, resulting in 11 sets of image data, to discuss the robustness of the maize LAI estimation model under different flight experiments. A multi-source remote sensing image dataset was constructed to provide image feature variables highly correlated with maize LAI. Eight texture information categories were generated based on gray-level co-occurrence matrix from the original image texture features. In the end, 51, 43, and 9 image features were obtained from RGB, MS, and TIR image data sources, respectively. Seven machine learning models, including GBDT, LightGBM, MLPR, PLSR, RFR, SVR, and XGBoost, were selected to estimate the vertical distribution of maize LAI. These models were applied to estimate LAI vertical distribution data at different maize growth stages. Two models with the strongest robustness were selected to verify the optimal observation time and flight altitude under different drone flight heights and sun elevation angles. The research results showed that during the reproductive growth stage of maize, the best single growth period for estimating maize LAI was the tasseling period. The MLPR model had R2 of 0.91 and rRMSE of 5.1% for LAI estimation. At the same time, the LAI estimation accuracy obtained during the maize maturation period was the worst, with R2 of 0.8 and rRMSE of 11.01%. As the measurement height of maize LAI was increased, the accuracy trend differred from that at the bottom, showing a trend of first decreasing and then increasing. Based on the experiments conducted involving different flight and solar altitude angles, it was concluded that lower flight altitudes of UAVs led to higher accuracy in estimating maize LAI. Specifically, at a flight altitude of 30m, the MLPR model achieved an accuracy of R2 of 0.73 and RMSE of 10.97%. Additionally, the highest accuracy in maize LAI observation was achieved when observations were conducted at 09:00 and 10:00 in the morning. The use of UAV remote sensing technology, combined with multi-source image data, enabled accurate observation of the vertical distribution of LAI in maize canopies. This approach enabled a precise understanding of the spatial distribution of maize LAI at different heights, and provided timely information on the health status of functional leaves. The acquired data can be used to adjust field management measures accordingly. Furthermore, experts in maize breeding can use this technology to identify differences between maize varieties and select specific cultivars, which had significant practical implications.