Accurate crop yield forecasting is critical for securing food supplies, addressing climate-related risks, and enhancing agricultural production planning. Current predictive models often have drawbacks, including cumbersome architectures, heavy computational load, and insufficient feature mining. To resolve these issues, this study develops a lightweight prediction model designed to reduce structural complexity and computational load while maintaining high accuracy based on a symmetric dual-branch attention mechanism. This model adopts a two-branch symmetric structure: the spatial branch processes remote sensing images and geospatial data via convolutional neural networks to capture crop growth-related spatial patterns effectively. The temporal branch analyzes meteorological and yield time-series data using long short-term memory (LSTM) networks to capture temporal variation trends precisely. The outputs of the two branches are deeply integrated through feature concatenation and an adaptive weighting strategy. To test the model’s performance, this study uses county-level yield records, long-term time series, and meteorological datasets from 1980 to 2018 from major U.S. soybean-producing states as experimental inputs. This dataset is then compared with leading models like AMAP. Results show that in single-year forecasting, the model reduces RMSE by 4.1762% and boosts R2 by 3.4458%—demonstrating strong short-term prediction capability. For five-year long-term forecasting, it reduces the RMSE by 3.2914% and increases the R2 by 4.7537%. This effectively mitigates the performance decline of traditional models in long time-series scenarios, fully leveraging the value of in-depth mining for long-term time-series data.
Li et al. (Mon,) studied this question.