Agriculture is the pillar of India's economy and food security, yet accurate crop yield prediction remains a persistent challenge due to the complex interplay of environmental, agronomic, and socio-economic factors. In practice, districts often share crop-related characteristics with more than just their immediate neighbours and exhibit irregular temporal patterns in variables such as rainfall. Existing models often fail to capture higher-order spatial dependencies between crop-producing regions and tend to overlook the irregular timing of key agricultural attributes, leading to reduced predictive accuracy. This study presents a Spatial Graph Hop with Temporal Enhancement (SGHTE) model that leverages multi-region spatial dependencies via an extended MixHop GCN and models irregular temporal patterns with a time-aware LSTM. By capturing cross-district interactions and long-range spatial–temporal correlations, SGHTE significantly improves crop yield prediction accuracy across Rajasthan’s 32 districts. The approach is further strengthened by incorporating a comprehensive set of 15 attributes, including underexplored factors such as saline and sodic soil composition, diverse irrigation sources, hybrid seed use, and fertilizer application, thereby enriching the feature space for prediction. Using a dataset from Rajasthan covering 32 districts over 13 years (2007–2019), SGHTE achieved an RMSE of 0.1306, an R² of 0.6775, and a Pearson correlation coefficient of 0.8912 for Bajra (Pearl Millet) yield prediction. These results demonstrate clear improvements over state-of-the-art methods, highlighting SGHTE’s capability to model complex spatial-temporal dependencies and provide reliable, actionable forecasts to support decision-making for both policymakers and farmers.