QuantumEarthNet a hybrid spatio-temporal deep learning framework for multi-horizon earth system forecasting

These require a proper forecast of Earth system dynamics to monitor climate, support environmental management, and enhance disaster preparedness. Nonetheless, predicting important climate variables over extended periods in the future is quite difficult because of the complex nonlinear processes within atmospheric, oceanic, and land systems. The classical statistical models, including ARIMA and SARIMA, use linear assumptions and thus cannot capture high-dimensional spatio-temporal dependencies. Recent deep learning networks, such as CNNLSTM, ConvLSTM, and transformer architectures, have enabled higher predictive power. Still, issues include long-range spatial dependencies, promoting spatial smoothness, minimising artefacts, and minimising error propagation through iterative prediction. To overcome these shortcomings, we introduce QuantumEarthNet, a hybrid spatio-temporal forecasting model that combines deep learning with quantum-assisted refinement of latent representations. To enhance physical plausibility, we propose a spatial smoothness and consistency regularisation term to reduce unrealistic spatial variations in the predicted climate fields. The architecture is implemented as a multi-scale temporal learning system, using a graph-based, teleconnecting model of space that represents the complex interactions among spaces. It presents a Quantum Latent Refinement (QLR) module that extends nonlinear feature-cell transformations within the compressed latent representations, and regularisation measures that foster spatial harmony and consistency in forecasting. Comparative analysis. It is shown that the suggested framework achieves improvements similar to those of baseline models, such as CNN-LSTM, Conv-LM, transformer, and graph-based models, when dealing with long forecasting horizons. These improvements are robust through statistical validation. The findings suggest that the hybrid design contributes to enhanced predictive accuracy and minimised multi-horizon predictive error, while incurring manageable computational overhead. On the whole, the work demonstrates the opportunities offered by hybrid quantum–classical architectures to advance near-term Earth system forecasting and climate intelligence applications.