Try GeoDiff-SR
Pick one of the bundled low-resolution UAV plot frames below to run it through the model.
Loading sample gallery…
18 real low-resolution UAV frames (9 sweet corn and 9 green beans) are available to test. Custom image upload is planned but not enabled yet.
Overview
High-resolution aerial imagery underpins precision agriculture, but the spatial resolution of UAV surveys is limited by sensor, altitude, and flight-time constraints. Computational super-resolution offers a software-side remedy, yet the GAN-based methods commonly used optimize image-fidelity metrics and are prone to hallucination, fabricating high-frequency content that can corrupt agronomic information and mislead decisions.
GeoDiff-SR is a conditional denoising-diffusion model that enhances the resolution of agricultural images by a factor of four while keeping reconstructions faithful to the underlying scene. It was trained on 2,136 sweet corn and green bean images collected over three growing seasons (2020 to 2023) at the UF/IFAS Tropical Research and Education Center, with low-resolution counterparts synthesized using a two-stage stochastic Real-ESRGAN-style degradation operator.
Method
Diffusion backbone
A conditional denoising-diffusion model learns to reconstruct the 4x high-resolution image, sampled with 50 DDIM steps using EMA weights.
Tiled sampling
64x64 low-resolution tiles with 16 px overlap are super-resolved to 256x256 and merged with Hann-window blending, scaling from a single patch to full orthomosaics.
Agronomic evaluation
Beyond PSNR, SSIM, and LPIPS, the model is judged on agronomic metrics (ExG-segmentation IoU, canopy-coverage error) and a human opinion-score study.
Results
GeoDiff-SR was compared against Real-ESRGAN and bicubic interpolation using image-fidelity, agronomic, and human-perception measures.
| Method | Mean Opinion Score ↑ | Canopy-coverage error ↓ |
|---|---|---|
| Bicubic | 0.45 | n/a |
| Real-ESRGAN | 0.82 | 2.37% |
| GeoDiff-SR | 0.98 | 0.71% |
GeoDiff-SR attained the highest mean opinion score (0.98, versus 0.82 for Real-ESRGAN and 0.45 for bicubic), increased ExG-segmentation IoU by 25%, and reduced canopy-coverage error from 2.37% to 0.71% relative to Real-ESRGAN. The advantage held on independent fixed-wing airplane imagery from the Florida Department of Transportation, demonstrating cross-platform generalization.
Human Perception Study
The mean opinion scores above come from a blind human-in-the-loop survey of 38 participants who rated reconstructions from each method. The interactive evaluation tool used for that study is hosted alongside this project.
Open the evaluation appCode & data
The trained model, inference code, and dataset are available upon request.