Unlocking High-Resolution Digital Soil Mapping With Hybrid Quantum–classical Deep Learning

Digital Soil Mapping (DSM) has become an essential tool for sustainable land management, enabling accurate prediction of soil properties over large areas. Algeria, characterized by diverse climatic zones and heterogeneous landscapes, presents both challenges and opportunities for DSM. In this study, we successfully generated detailed Algerian soil maps of 700m spatial resolution by integrating soil observations with covariates derived from the SCORPAN framework to develop deep learning–based models for soil prediction. A multi-task convolutional neural network (MTCNN) and a hybrid quantum–classical MTCNN (HQC-MTCNN) were implemented. The dataset was carefully prepared to serve as input–output pairs for the models, ensuring compatibility with convolutional architectures. Results demonstrate that the classical MTCNN consistently outperforms the HQC-MTCNN, achieving lower mean absolute error (MAE) and root mean squared error (RMSE) across all tasks. These findings highlight the robustness of deep learning approaches for DSM in data-limited regions and underline the current limitations of quantum–classical models when simulated on classical hardware. Nevertheless, the study provides valuable insights into the integration of emerging quantum techniques with soil science and lays the groundwork for future exploration once quantum hardware matures.