UDAMSR Net: An Unsupervised Degradation-Aware Network for Enhancing the Spatial Resolution of Spectral Images for Crop Sensing

Low spatial resolution (LR) remote sensing data is widely adopted because of its lower cost, although its limited analytical precision constrains its full use in precision agriculture. By contrast, the acquisition of high spatial resolution (HR) data often requires substantial expense. To address this limitation, this study proposes an unsupervised degradation-aware multi-channel super-resolution network (UDAMSR) to enhance LR spectral images without requiring paired HR–LR training data. The main contributions are as follows: ① the original framework is extended with dedicated queue and reconstruction layers to process multispectral and hyperspectral image (HIS) cubes, and a contrast-learning-based degradation-aware module is integrated to address unknown real-world degradation; ② comprehensive evaluation is conducted using image quality metrics, spectral consistency analysis, and performance in crop remote sensing tasks, such as chlorophyll content estimation; ③ the generalization capability of the model is assessed using data from three imaging devices, two spatial scales (near-ground and unmanned aerial vehicle (UAV)), and two geographic regions. The results show that the proposed method achieves the best overall performance in the comprehensive evaluation, with a mean peak signal-to-noise ratio ( P S N R ¯ ) of 32.78, a mean root mean squared error ( R M S E ¯ ) of 6.93, a mean structural similarity index ( S S I M ¯ ) of 0.89, and a mean spectral angle mapper ( S A M ¯ ) of 0.131. The method effectively reduces the degradation in chlorophyll detection accuracy caused by spatial resolution reduction. The evaluation of generalization capability further shows that the proposed method demonstrates strong generalization across different spatial scales, geographic regions, devices, and data types. These results indicate that UDAMSR provides a robust, efficient, and cost-effective software solution that can compensate for hardware limitations and support high-quality crop phenotyping detection in diverse application scenarios.