关键词: 叶面积指数, 无人机遥感, 机器学习, 尺度转换

Abstract:

By comprehensively comparing the simulation accuracy, spatial distribution, and data distribution histograms of four machine learning algorithms (Lasso regression, Ridge regression, Gaussian process regression, and Random Forest regression), on the Leaf area index (LAI) of winter wheat under various feature combinations, a suitable unmanned aerial vehicle model for monitoring the LAI of winter wheat in the north China region was selected. Using the monitoring results from this model as ground truth, a Sentinel−2 winter wheat LAI monitoring model was developed to dynamically monitor and evaluate the LAI of winter wheat in Hebi city, based on the distribution of cultivated land. This study addressed the scale−up challenge in satellite remote sensing leaf area index (LAI) modeling, by innovatively introducing multispectral Unmanned aerial vehicles (UAV) as an intermediate scale. The results showed that: (1) among the four machine learning algorithms applied to UAV multispectral data, Lasso regression achieved the highest simulation accuracy (RMSE=1.472), followed by Ridge regression (RMSE=1.488), Gaussian process regression (RMSE=1.538) and Random forest regression (RMSE=1.582). The Ridge regression provided a balanced performance in both high and low values, while Random forest regression overestimates low values while underestimates high values, Lasso regression tended to overestimate low values and Gaussian process regression underestimated both extremes. The result histograms for Gaussian process regression, Lasso regression and Ridge regression exhibited a normal distribution, however, the histogram of Random forest regression displayed greater dispersion. Consequently, Ridge regression utilizing 18 features was confirmed to be the optimal model for monitoring LAI using UAV. (2) For Sentinel−2 based modeling, the algorithm performance ranked as Ridge regression > Lasso regression > Gaussian process regression > Random forest regression, and the Ridge regression utilizing 26 features was confirmed to be the optimal model for Sentinel−2 LAI monitoring. (3) The dynamic monitoring of cropland LAI in Hebi using Sentinel−2 data revealed that the average LAI values on March 28, April 27 and May 12 were 2.50, 3.22 and 2.92, respectively. This demonstrated stable monitoring with a higher spatial resolution than MODIS product.

Key words: Leaf area index (LAI), Unmanned aerial vehicles remote sensing, Machine learning, Scale transition