Winter wheat yield estimation based on multisource remote sensing data: A dual-branch TCN-Transformer model and analysis of growth-stage feature transition mechanisms

Abstract

Timely and accurate acquisition of winter wheat yield information is crucial for ensuring food security and formulating agricultural policies. Although deep learning methods have become increasingly prominent in crop yield estimation, they often face challenges in simultaneously capturing both fine-grained local patterns and long-term temporal dependencies in time series data. By utilizing EVI, LAI, and fraction of photosynthetically active radiation (FPAR) from MODIS, along with temperature (TEM) and precipitation (PRE) data from ERA5-Land, we propose a novel dual-branch hybrid model named TCN–Transformer (TCT), which synergistically integrates temporal convolutional network (TCN) and transformer architectures to concurrently capture both localized temporal patterns and long-term dependencies. Bayesian optimization was employed for automated hyperparameter tuning, enabling accurate estimation of winter wheat yield under diverse agricultural management conditions. The experimental results demonstrate that optimal performance is achieved by the proposed TCT model in terms of estimating the county-level winter wheat yields across North China on the test set (R2 = 0.80, RMSE = 645.75 kg/ha). It significantly outperforms the individual temporal models (the TCN, LSTM, and transformer) and other comparative models, including traditional machine learning methods (Ridge, RF, LightGBM, and XGBoost) and an advanced hybrid model (CNN-BiLSTM). Specifically, compared with the individual models, the TCT improved R2 by 0.03 to 0.1 and reduced the RMSE by 29.33 to 156.07 kg/ha. It also outperforms CNN-BiLSTM (R2 = 0.78, RMSE = 668.23 kg/ha), achieving lower errors and more robust bias control. To elucidate the decision-making mechanism of the model, the Shapley additive explanations (SHAP) method was employed to analyze the feature importance values across the study region and the temporal feature weights at 8-day intervals. The results reveal that the EVI is the most representative feature, with the model accurately identifying critical growth stages from T20 (February 26) to T28 (May 1), corresponding to the greening to milk phases, respectively. The feature contribution dynamics were further visualized, revealing a transition from FPAR dominance during early greening (T20–T22) to EVI dominance during jointing (T23–T25), EVI‒PRE interactions during heading-milk (T26–T29), and finally LAI‒PRE dominance at maturity (T30–T32). Furthermore, the one-year leave-one-out cross-validation confirms the robustness of the TCT model, the simulation of yield spatial distribution for unseen years is consistent with the official yield data. Additionally, the proposed interpretability framework not only performed excellently in this study but also demonstrated strong generalizability and flexibility, indicating its broad application potential in other crop types and agricultural domains.