Feature Fusion for Winter Wheat SPAD Estimation Using Pooled Samples from Multiple Growth Stages: A Multi-Source Deep Learning Approach

SPAD values indicate relative leaf chlorophyll content and support winter wheat growth and nitrogen monitoring. This study evaluated spectral–texture feature fusion using pooled samples from multiple growth stages. UAV multispectral imagery and 288 SPAD samples were collected at jointing, booting, early and late grain filling in 2021 and 2022. Vegetation indices (VIs) and texture features (TFs) were calculated from the imagery. Multidimensional spectral feature combination indices (D-SFCIs) and multidimensional texture feature combination indices (D-TFCIs) were generated using combination formulas from previous studies, with 12 features selected for each group. A multi-source deep feature neural network (MSDNN) used separate fully connected branches for spectral and texture feature extraction, followed by deep feature fusion and joint training. It was compared with partial least squares regression (PLSR), random forest (RF), and a deep neural network (DNN). In the sample-level split, MSDNN achieved the highest accuracy among the four models using the same 48 spectral and texture features (R2 = 0.866, RMSE = 5.222, RRMSE = 14.9%). Feature screening preceded this split, and some plots contributed observations to both the training and test sets. In the primary evaluation on held-out plots, MSDNN achieved a lower test RMSE than DNN in one of three splits. Across the three splits, MSDNN performed similarly to DNN but did not outperform it on average (mean R2: 0.796 vs. 0.802, RMSE: 6.424 vs. 6.290, RRMSE: 17.8% vs. 17.4%). Because this evaluation used only fused inputs, the benefit of fusion over spectral or texture features alone remains to be tested on held-out plots.