Evaluating Crop Yield Sensitivity to Climate Change Using a Deep‐Learning Framework Incorporating Vegetation and Groundwater Memory

Summary

This paper presents an interpretable spatiotemporal deep-learning framework to model crop yield sensitivity to climate change in Northeast China, overcoming the temporal mismatch between satellite-era observations and multi-decadal yield records by reconstructing vegetation and groundwater memory back to 1981. The GATAE-TFT model identified vegetation greenness and hydro-climatic memory as the most influential predictors, and climate-scenario projections (2017–2060) indicate gradual declines in maize, rice, and wheat yields whilst soybean yields remain stable, providing a physically grounded benchmark for assessing water-stress risks to grain production under future climate scenarios.

Regional applicability

This study is specific to Northeast China's agroclimate and grain production systems. Findings may have limited direct transferability to United Kingdom conditions, which differ markedly in climate regime, crop types (wheat and barley dominate), groundwater dynamics, and seasonal patterns. However, the methodological approach—integrating vegetation and hydrological memory into yield models—could inform UK research on climate-resilient crop forecasting, particularly for water-stress assessment in drier regions of eastern England.

Key measures

Crop yield (maize, rice, soybean, wheat); Normalized Difference Vegetation Index (NDVI); groundwater storage anomalies; drought indices; model R² values; SHAP-based feature attribution

Outcomes reported

The study developed a spatiotemporal deep-learning model (GATAE-TFT) to reconstruct historical crop yield responses to weather, vegetation, and groundwater memory from 1981 onwards, and projected future yield changes under climate scenarios to 2060. Model validation achieved test R² values of 0.857–0.837 across maize, rice, soybean, and wheat.

Topic tags