TDA-Net: Two-Branch Deformable and Atrous Network for Estimating Aboveground Biomass in Rubber Plantations

The aboveground biomass (AGB) is an important index for assessing the growth of crop rubber trees and is directly related to the assessment of the health status and carbon storage of rubber plantations. Remote sensing imagery combined with deep learning (DL) is a powerful tool for estimating crop AGB. However, the current DL models are limited to using a convolution kernel with a fixed geometry structure and a single channel to extract the spatial and spectral features of multispectral images. Therefore, the accuracy of the existing AGB estimation models needs further improvement. In this study, we propose a two-branch deformable and atrous network (TDA-Net) for AGB estimation. We construct two branches of convolutional neural networks: a deformable convolutional module to extract the spatial features and an atrous convolutional module to extract the spectral features from satellite images. The AGB of rubber plantations is estimated using the proposed TDA-Net. The experimental results show that the estimation performance of the proposed TDA-Net is superior ( ${R}^{2}$ = 0.802) to that of the currently used models. The results of this study can provide a basis for crop AGB estimation. To the best of our knowledge, this is the first time using submeter-resolution satellite images for the AGB estimation.