A Novel Hybrid-DCNN-Based Framework for Enhanced Rice Aboveground Biomass Estimation Under Limited Samples

Aboveground biomass (AGB) of rice is crucial for monitoring growth and predicting yields. While deep learning algorithms, such as deep convolutional neural networks (DCNNs), show compelling performance in estimating crop parameters, gathering sufficient ground-truth samples for model training poses a significant challenge, leading to the “small sample problem.” To address this, we propose a framework that utilizes a hybrid inversion model based on the PROSAIL-PRO radiative transfer model (RTM) combined with machine learning techniques [XGBoost and random forest (RF)]. This framework incorporates active learning optimization and the spectral angle mapper (SAM) method to select simulated samples that closely match real-world conditions, simultaneously assigning geographic location information to the samples. Using these qualified samples, we constructed both single-branch and multibranch DCNN models that integrate uncrewed aerial vehicle (UAV)-based hyperspectral principal components (PCs), canopy height (CH) information from the canopy surface model (CSM), and canopy temperature derived from thermal infrared (TIR) images. The effectiveness of this approach was validated across two experimental sites. The single-branch DCNN achieved the highest accuracy at site A ( $R^{2} =0.816$ and root-mean-square error (RMSE) =61.608 g/m2) with PCs, TIR, and CSM as inputs, while the multibranch DCNN performed best at site B ( $R^{2} =0.784$ and RMSE =65.533 g/m2), using PCs and TIR as inputs. Results indicate that simulated samples have considerable potential for practical applications. PCs were the primary contributors to the model, with TIR playing a more significant role than CSM. Overall, this study demonstrates high-precision estimation of rice AGB despite limited measured samples, offering valuable insights for crop monitoring under small sample conditions.