Application of U-net models in estimating forest canopy closure based on multi-source remote sensing imagery

Forest Canopy Closure (CC) is vital for assessing forest ecosystems. This study integrates multispectral imagery with enhanced U-Net models (U-Net, U-Net++, U-Net3+) to achieve cost-effective large-scale CC estimation. These models are optimized by reordering the network output layers and enhancing feature fusion between convolutional and pooling operations. By experimenting with different combinations of multi-parameters with the improved U-Net architectures, we estimate CC and validate the results using airborne Light Detection and Ranging (LiDAR) CC data. The results show that (1) Ratio Vegetation Index (RVI) had the strongest correlation with CC (R2= 0.8135). (2) U-Net exhibits optimal stability under structural adjustments; (3) Adjust-U-Net++ achieves the highest accuracy, with R2=0.8785, RMSE = 0.1256, EA = 77.46% and MAE = 0.0800; (4) Multi-parameter combinations outperform single parameters in CC estimation. By exploring both the selection of input parameters and the structural optimization of U-Net models, this study provides an effective approach for large-scale, low-cost CC estimation.