End-to-End Optimized DOE for Spatial Classification in Agricultural Imaging

Precision agriculture aims to optimize crop production using advanced sensing and computational methods. Traditional RGB imaging systems often fail to capture the subtle spectral differences necessary for accurate classification of materials in agricultural scenes, as many crop and soil properties exhibit unique spectral signatures in the visible range. This paper introduces a compact, static optical system based on a diffractive optical element (DOE), jointly optimized with a spatial classification model through an end-to-end (E2E) learning framework. The goal is to enhance spatial classification performance without increasing system complexity. This work models light propagation using angular spectrum theory, employ Zernike polynomial parameterization to constrain the DOE design space, and integrate a neural network trained on CALCROP21 data. Simulations demonstrate that the proposed E2E-optimized DOE outperforms baseline RGB-based systems in classification accuracy while maintaining high light efficiency and enabling offline processing. These findings highlight the potential of task-specific DOE design for embedded agricultural imaging systems, achieving an accuracy increase in the best cases up to 20.9% compared to conventional RGB-based systems.