A research team from China Agricultural University and partner institutions has introduced an unsupervised neural network to upgrade low-spatial-resolution spectral remote sensing data for crop monitoring, with findings published in Engineering.

Low-spatial-resolution remote sensing imagery remains widely adopted due to its low acquisition cost, yet its limited spatial detail weakens analytical precision for core crop phenotyping tasks such as chlorophyll content estimation. High-spatial-resolution spectral data, by contrast, carries substantial hardware and field operation costs, creating a persistent trade-off between spatial detail, spectral band count and sensor expense.

Traditional resolution enhancement methods face two key practical barriers: pan-sharpening algorithms require matched high-resolution panchromatic images that are rarely available in field settings, while mainstream supervised single-image super-resolution models rely on scarce paired high-resolution and low-resolution training samples and fail to account for complex, unknown real-world imaging degradation.

The team’s proposed framework, named UDAMSR Net, builds on the existing DASR blind super-resolution architecture with targeted modifications tailored to multispectral and hyperspectral image cubes.

The network integrates two core functional modules: a contrastive learning-driven degradation-aware module that extracts compact degradation representation vectors to characterize diverse real-world image distortions, and newly designed queuing and reconstruction layers that split multi-channel spectral cubes into single-channel grayscale sequences for processing and reassemble enhanced outputs back into complete spectral data cubes.

A dual-branch degradation attention block inside the super-resolution backbone injects degradation feature information to modulate spatial and channel-wise features simultaneously, eliminating visible mesh artifacts observed when applying the original DASR model to multispectral agricultural data.

To mitigate the shortage of agricultural remote sensing training datasets, the research adopts transfer learning, first pretraining the network on combined DIV2K and Flickr2K natural image datasets converted to grayscale, then fine-tuning the model with a field-collected transfer learning dataset built from wheat canopy hyperspectral data.

Field validation was carried out using winter wheat datasets gathered at two experimental sites in Hengshui, Hebei Province and Xinxiang, Henan Province, covering three imaging devices including near-ground hyperspectral cameras, near-ground multispectral cameras and UAV-mounted multispectral sensors across near-ground and aerial spatial scales.

Model performance was measured using spatial image metrics including PSNR, RMSE and SSIM alongside the spectral angle mapper (SAM) to quantify spectral consistency between reconstructed outputs and original high-resolution imagery.

The transfer learning-adjusted T-UDAMSR variant outperformed bicubic interpolation and the pre-trained P-UDAMSR model across all tests, delivering consistent restoration of leaf edge textures and canopy structural features.

When deployed for downstream chlorophyll content prediction via partial least squares regression (PLSR) modeling, imagery processed by T-UDAMSR restored estimation accuracy degraded by low spatial resolution, with stable predictive performance maintained across different sensors, flight altitudes and geographic field sites.

Texture feature analysis further confirmed that super-resolution processing increased angular second moment and homogeneity while lowering entropy and contrast, producing more regular, noise-suppressed canopy textures.

The study frames UDAMSR as a cost-efficient software alternative to costly hardware upgrades for agricultural spectral sensing, and outlines four directions for follow-up work including cross-crop generalization testing and refined canopy segmentation strategies to advance field-scale practical deployment.

The paper “UDAMSR Net: An Unsupervised Degradation-Aware Network for Enhancing the Spatial Resolution of Spectral Images for Crop Sensing,” is authored by Weijie Tang, Ruomei Zhao, Hong Sun, Minzan Li, Lang Qiao, Mingjia Liu, Guohui Liu, Yang Liu, Di Song. Full text of the open access paper: https://doi.org/10.1016/j.eng.2026.01.031. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.