Hybrid Quantum-Classical Deep Learning Model for Global Cotton Yield Forecasting

Abstract

Global cotton yield forecasting is critical for food security planning, supply chain management, and climate adaptation policy, yet remains challenging due to the non-linear interactions among satellite vegetation indices, meteorological variables, and agronomic factors across heterogeneous growing regions. This paper proposes the Hybrid Quantum-Classical Deep Learning (HQC-DL) framework, the first systematic empirical study integrating parametric quantum circuit (PQC)-based feature enhancement into a CNN-BiLSTM and gradient-boosted regression pipeline for global crop yield prediction. When tested on the FAOSTAT-MODIS-CRU dataset with 1,142 country-year observations of 59 countries between 2001 and 2020, the proposed method exhibits RMSE of 0.769±0.004 t/ha and R2 of 0.690± 0.004, and is characterized by the lowest cross-run variance compared to all other neural network architectures, which demonstrates its better prediction stability. The ablation study shows that all architectural stages have a positive contribution: when gradient boosted regressors are trained on features generated via deep learning methods, they show better results than models trained on raw data, regardless of the quantum module, and PQC expressibility analysis proves that the ansatz is capable of exploring a wide range of states in Hilbert space.