Efficient hybrid quantum-classical neural networks generation for crop yield prediction using sequential satellite images

Abstract A novel approach is introduced that enables the automatic generation of hybrid quantum-classical neural networks using a two-step optimization process, combining multiobjective genetic algorithms specifically designed for time series on satellite images with a subsequent continuous refinement stage of the circuit’s parameters. This strategy provides models with low prediction error and low complexity in terms of trainable parameters. The application of this methodology focuses on using Sentinel-2 satellite images to forecast future values of the Normalized Difference Vegetation Index (NDVI), which holds great significance in agriculture since it is used as an indicator of crop health and yield. The hybrid model is evaluated on random validation pixels and full images, and compared against tuned classical baselines to assess its performance.