MHISR: A spatial-spectral fusion network for microscopic hyperspectral image super-resolution with application in rice disease evaluation

Stomata are the primary microscopic entry points for Xanthomonas oryzae pv. oryzae ( Xoo ) invasion, serving as key physiological indicators for early and accurate disease diagnosis. Although microscopic hyperspectral imaging (MHI) captures microstructural and physiological alterations in plants, hardware-induced spatial resolution limitations constrain its efficacy. To overcome these constraints, this study proposes MHISR, a deep learning-based super-resolution architecture to enhance spatial resolution while preserving spectral information. The architecture employs a bidirectional Mamba module to capture long-range dependencies in the state space during spectral processing and further incorporates the residual transposed attention mechanism (Res-MDTA) to enhance global perception of spatial textures. Experimental results based on microscopic datasets of four crops (rice, maize, tomato, and chili pepper) demonstrate that MHISR outperforms state-of-the-art (SOTA) models in image quantification metrics and visual quality. Specifically, the model achieved the highest peak signal-to-noise ratio (PSNR, 37.33 dB), with spectral angle mapper (SAM) and relative global dimensionless synthesis error (ERGAS) improved by 4.03% and 15.22% over the second-best model. Furthermore, a classification model trained on the super-resolution (SR) data achieved an accuracy of 0.96 (F1-score = 0.94). Based on the spectral and textural features of the SR data, a rice stomatal disease resistance index (SDRI) was developed to quantify disease resistance at the microscopic scale. MHISR reduces classification errors caused by blurred microscopic details, providing an advanced method for the high-throughput screening of rice varieties with disease resistance.