Abstract:
Leaf Area Index (LAI) can often be used to represent the growth status of vegetation. The absorption ratio of crop Photosynthetically Active Radiation (PAR) can then be determined to monitor the vegetation growth and crop yield estimation. Taking the mangrove as the research object, this study aims to realize the field measurement for the rapid and accurate estimation of large-scale LAI. The multi-spectral images were collected to serve as the data source using the Unmanned Aerial Vehicle (UAV) and Sentinel-2A (S2) in the North Gulf of Guangxi in South China. After that, the original spectral bands of UAV and S2 were used to calculate 16 vegetation indices and their combinations under various images. High-dimensional datasets of different mangrove species were constructed to integrate the original bands, vegetation indices, and combined vegetation indices of different images with the ground-truth LAI data. Some operations were then selected to process the datasets for the better normality of the data and the higher estimation accuracy of LAI models, including the data normalization, Box-Cox transformation, logarithmic transformation, model fitting, and the removal of outliers. The maximum correlation coefficient method was used to remove the redundant bands for the low dimensionality of the high dimensional datasets. The characteristic bands were then obtained to estimate the LAI of different mangrove species. Subsequently, the feature optimization was performed on the characteristic bands of forest species. The optimal feature dataset was established for the LAI estimation of mangrove tree species. Six machine learning algorithms (XGBoost, Feedforward Back Propagation (BP), Support Vector Machine (SVM), Ridge Regression (Ridge), Lasso) and ElasticNet were selected for the preferred feature bands of LAI of different mangrove species, according to the optimal characteristic band dataset. Model training and parameter tuning were also carried out for the different models at a ratio of 7:3 for the training set and test set. The performances of 6 machine learning algorithms were evaluated for the LAI of mangrove species. The accuracy of LAI estimation on the mangrove species was compared to determine the optimal models and image combinations for the various tree species. The research results indicated that: 1) The XGBoost model achieved the highest precision estimation of mangrove LAI under UAV and S2 image data, where the coefficient of determination (R2) was higher than 0.70, and the Root Mean Square Error (RMSE) was lower than 0.349. 2) The R2 values of the XGBoost model were improved by 0.105-0.365 and 0.283-0.540, respectively, and the RMSEs were reduced by 0.100-0.392 and 0.102-0.518, respectively, compared with the other five models. 3) The XGBoost mangrove LAI model that was constructed by UAV image data performed a better estimation accuracy for the Avicennia marina (AM) LAI than the rest images and combinations (R2=0.821, RMSE=0.288). The XGBoost mangrove LAI model constructed by S2 image data presented the excellent estimation accuracy (R2=0.940-0.979, RMSE=0.104-0.142) for the LAI of Kandelia candel (KC) and Aegiceras corniculatum (AC). The R2 values of LAI models for the different mangrove tree species under UAV and S2 images were ranked in the order of the Aegiceras corniculatum > Kandelia candel > Avicennia marina. 4) The R2 values of the S2 image for the mangrove LAI were better than those of UAV images. The SNAP-SL2P underestimated the mangrove LAI values as a whole. The UAV images presented a better estimation accuracy for the mangrove species LAI than the S2 images.