MultiAltitudeSR: Multi-Altitude Agricultural UAV Image Super-Resolution via Physical Ground Sampling Distance Conditioning

Xintong Jiang; Yu Tian; Yixue Liu; Valerio Hoyos-Villegas; Viacheslav Adamchuk; Shangpeng Sun

cs.CVOct 18, 2026

MultiAltitudeSR: Multi-Altitude Agricultural UAV Image Super-Resolution via Physical Ground Sampling Distance Conditioning

Authors: Xintong Jiang, Yu Tian, Yixue Liu, Valerio Hoyos-Villegas, Viacheslav Adamchuk, Shangpeng Sun

Organizations: Department of Bioresource Engineering, McGill University, 21111 Lakeshore Road, Ste-Anne-de-Bellevue, H9X 3V9, Quebec, Canada · Michigan State University, 426 Auditorium Road, East Lansing, 48824, Michigan, United States

Abstract

Unmanned aerial vehicles (UAVs) enable high-throughput platforms for crop monitoring but face a fundamental trade-off between flight altitude and image spatial resolution. Many existing super-resolution (SR) methods for UAV imagery rely on oversimplified synthetic degradation models and are restricted to fixed integer upsampling scales, limiting their ability to handle real-world degradations and varying flight altitudes. This study proposes MultiAltitudeSR, a framework that conditions feature modulation and reconstruction on source and target ground sampling distances (GSDs). MultiAltitudeSR trained jointly on real UAV images at altitudes of 10 to 70 m above ground level supports all 21 source-to-target altitude pairs without pair-specific fine-tuning, covering scale factors ranging from 1.18 $\times$ to 6.03 $\times$ . The framework can be extended to additional source-target GSD combinations by incorporating corresponding image pairs into training. On the 20-to-10 m pair, the model achieved competitive perceptual performance compared with pair-specific baselines, with LPIPS of 0.2236, DISTS of 0.2431, and NIQE of 8.9020. Experiments using the same architecture demonstrated the value of acquiring real-world multi-altitude UAV image pairs for SR training, with real-pair training achieving better average reconstruction performance across the 21 altitude pairs than training with the tested synthetic degradation pipeline. Downstream zero-shot segmentation case studies illustrated task-dependent benefits of super-resolved imagery for crop monitoring, suggesting potential UAV flight-time savings through image acquisition at higher altitudes.

Figures & tables

Figure 1: Overview of the multi-altitude UAV dataset that consists of multiple crop and weed species.

Figure 2: Comparison of image details and luminance surfaces across flight altitudes for common bean and chickpea, two crops with contrasting leaf sizes.

Figure 3: OTF synthetic LR dataset. (a) The OTF synthetic LR generation process, in which only real-world 10 m HR images are used. A random degradation is drawn from the degradation distribution pool every training epoch. Based on each real-world 10 m HR image, a fresh LR input is synthesized by the high-order degradation inside the data loader every epoch during training time. Intermediate GT targets from 20 m to 70 m are bicubic downsamples of the same HR, used as supervision targets during both training and validation. (b) Violin plots of three statistics measuring the difference between the synthetic and real-world datasets. (c) PSD ratio across spatial frequency for each altitude. (d) Fidelity between the real-world and synthetic LR distribution overlaps for different tested crops and altitudes.

Figure 4: Model overview. UAV imagery of the same region of interest (RoI) is acquired at seven altitudes AGL. An RRDB backbone, initialized with pretrained weights and fine-tuned during training, extracts image features. The content-aware encoder derives a spatial quality descriptor from LR inputs, while the GSD Fourier encoder embeds the source and target GSDs. Together, these encodings condition the modulation of backbone features through scale-conditioned FiLM. The GSD-conditioned reconstruction head applies GSD-conditioned filtering to resample features to the target resolution. Source-GSD-based hard routing then selects one of six output experts to reconstruct the super-resolved image.

Figure 5: GSD-conditioned adaptive reconstruction head. FiLM-modulated backbone features undergo shared refinement at the source resolution, GSD-adaptive upsampling, and shared refinement at the target resolution. Separate pre-filter and post-filter MLPs map the log source-to-target GSD ratio to eight group-shared 5×5 and 9×9 kernels, respectively, with each kernel normalized by softmax. These depthwise filters operate before and after bilinear resizing. Hard routing selects the expert whose GSD center is nearest to the source GSD. Within the selected expert, a linear projection of the joint conditioning vector $\mathbf{z}$ provides a channel-wise bias, added after the first convolution and before activation. The expert performs conditioned projection, feature refinement, and RGB residual prediction. The predicted RGB residual is added to the bicubic-resized input to produce the super-resolved output.

	Perceptual			Fidelity
Method	LPIPS $\downarrow$	DISTS $\downarrow$	NIQE $\downarrow$	PSNR (dB) $\uparrow$	SSIM $\uparrow$	MS-SSIM $\uparrow$
Bicubic	0.3961	0.2706	14.5626	15.84	0.2945	0.6499
EDSR	0.5393	0.3406	11.0497	16.83	0.3020	0.6631
SwinIR	0.2265	0.2503	7.7757	16.16	0.2747	0.6421
Real-ESRGAN	0.2144	0.2365	9.4331	15.78	0.2763	0.6473
Ours	0.2236	0.2431	8.9020	16.04	0.2684	0.6497

Table 1: Quantitative comparison with state-of-the-art methods on the 20

\rightarrow

10 m altitude pair. Bold values indicate the best performance in each column.

Figure 6: Qualitative comparison of 20 $\rightarrow$ 10 m super-resolution results across seven representative scenes. Each column shows the same tile as the HR reference and as reconstructed by bicubic interpolation and the different super-resolution models.

Figure 7: Zoomed-in comparison of 20 $\rightarrow$ 10 m super-resolution results for two representative scenes. The left panel shows the HR reference captured at 10 m AGL, with the red box marking the magnified region. The right panels show the corresponding region from the 20 m LR input and the outputs of bicubic interpolation, EDSR, Real-ESRGAN, SwinIR, and the proposed model.

Target	Source	Synthetic / Real
Target	Source	PSNR (dB) $\uparrow$	SSIM $\uparrow$	MS-SSIM $\uparrow$	MAE $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$
10 m	20 m	20.01 / 15.49	0.5944 / 0.2755	0.8803 / 0.6460	0.0811 / 0.1329	0.1826 / 0.2257	0.2035 / 0.2277
	30 m	19.17 / 14.99	0.5197 / 0.2159	0.8509 / 0.5957	0.0890 / 0.1412	0.2035 / 0.2703	0.2183 / 0.2597
	40 m	18.79 / 14.95	0.4893 / 0.1962	0.8340 / 0.5738	0.0927 / 0.1419	0.2118 / 0.3059	0.2249 / 0.2878
	50 m	18.42 / 14.98	0.4458 / 0.1808	0.8133 / 0.5530	0.0967 / 0.1424	0.2338 / 0.3318	0.2315 / 0.3001
	60 m	17.96 / 14.81	0.3986 / 0.1752	0.7881 / 0.5361	0.1017 / 0.1467	0.2568 / 0.3517	0.2424 / 0.3149

Table 2: SR evaluation of the synthesis-trained model, grouped by target altitude. Each cell reports the score on the synthetic / real test set; MAE is computed on intensities normalized to

[0,1]

, and Mean is computed over all 21 pairs. The lower panel reports the no-reference NIQE, which was computed only for the 10 m target and is therefore averaged over 6 pairs.

Figure 8: Perception-fidelity comparison on the 20 $\rightarrow$ 10 m pair. All tested SoTA models, the proposed model trained on real-world pairs (Ours), and the same architecture trained on the synthetic LR dataset (Synthesis LR) are placed on a perceptual-quality (LPIPS $\downarrow$ , DISTS $\downarrow$ ) versus fidelity (PSNR $\uparrow$ , SSIM $\uparrow$ ) plane, with both axes normalized to $[0,1]$ so that larger values indicate better quality. Representative SR results are shown beside each method and in the right column. EDSR achieves the highest PSNR and SSIM, and bicubic interpolation also scores relatively high on SSIM, yet both produce over-smoothed outputs, illustrating that higher fidelity scores do not necessarily correspond to better perceptual reconstruction.

Figure 9: Qualitative super-resolution results for seven representative source-to-target altitude pairs, spanning SR scale factors from $1.20\times$ (60 $\rightarrow$ 50 m) to $6.03\times$ (70 $\rightarrow$ 10 m). Each row shows three representative crops (left: nearest neighbor interpolation of low-resolution images; right: our SR outputs) drawn from distinct scene regions to capture varied canopy density and ground cover composition. GSD values (mm/pixel) for each pair are indicated on the left.

Target	Source	Scale	PSNR (dB) $\uparrow$	SSIM $\uparrow$	MS-SSIM $\uparrow$	MAE $\downarrow$	LPIPS $\downarrow$	DISTS $\downarrow$
10 m	20 m	1.67 $\times$	16.04	0.2684	0.6497	0.1263	0.2236	0.2431
	30 m	2.55 $\times$	15.62	0.2134	0.6012	0.1337	0.2864	0.2698
	40 m	3.42 $\times$	15.63	0.1978	0.5834	0.1337	0.3311	0.2863
	50 m	4.24 $\times$	15.64	0.1846	0.5635	0.1344	0.3485	0.3002
	60 m	5.09 $\times$	15.55	0.1810	0.5534	0.1365	0.3910	0.3114
	70 m	6.03 $\times$	15.56	0.1751	0.5415	0.1366	0.4315	0.3291

Table 3: SR evaluation of the proposed model trained on real-world data and evaluated on the real-world test set, grouped by target altitude. Bold : best value within each group. MAE is computed on intensities normalized to

[0,1]

, and Mean is computed over all 21 pairs. The lower panel reports the no-reference NIQE, which was computed only for the 10 m target and is therefore averaged over 6 pairs.

Panel of Table 5 — Table 5: Impact of wrong expert routing. For each pair, the correct GSD expert is replaced by the most distant available expert. $\Delta=\mathrm{mismatched}-\mathrm{correct~routing}$ . Negative values indicate degradation for PSNR, SSIM, and MS-SSIM, whereas positive values indicate degradation for MAE, LPIPS, and DISTS. Orange indicates improvement and blue indicates degradation, accounting for each metric’s direction.

Figure 10: Conditioned feature space visualized by crop type and altitudes. (a) Points colored by crop types. (b) Points colored by input altitudes. (c) Per-crop points colored by input altitudes.

Figure 11: GSD-calibrated patch-level LAM of three crops across all source altitudes targeting the same 10 m HR. Panels (a), (b), and (c) are representative LAMs of common bean, canola, and strawberry, respectively. The white cross marks the attribution centroid, and the cyan circle is the $\text{ERF}_{50}$ radius that encloses 50% of the attribution mass.

Figure 12: Zero-shot segmentation applications using SAM 3. (a) Weed detection in a common bean field at a middle-to-late growth stage, with SR from 40 $\rightarrow$ 20 m. (b) Crop-row segmentation of soybean grown with multiple cover crops, with SR from 60 $\rightarrow$ 30 m. (c) Detection of common bean seedlings at a very early growth stage, with SR from 10 $\rightarrow$ 5 m. Numbered boxes mark the zoom-in areas discussed in the text.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Panel of Table 17 — Table S1: Impact of average-case expert misrouting. For each altitude pair, the reported $\Delta$ is the mean degradation averaged over all five incorrect expert assignments. $\Delta=\mathrm{mismatched}-\mathrm{correctrouting}$ . Negative values indicate degradation for PSNR, SSIM, and MS-SSIM, whereas positive values indicate degradation for MAE, LPIPS, and DISTS. Orange indicates improvement and blue indicates degradation, accounting for each metric’s direction.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 248 papers

Altitude-Adaptive Vision-Only Geo-Localization for UAVs in GPS-Denied Environments
Feb 27, 2026Xingyu Shao, Mengfan He, Liangzheng Sun +2Cross-View Geo-Localization Map-Based Localization
Department of Precision Instrument, Tsinghua University, Beijing 100084, China · School of Instrumentation Science and Optoelectronics Engineering, Beijing Information Science and Technology University, Beijing 100192, China · School of Aerospace Engineering, Beijing Institute of Technology, Beijing 100081, China
TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution
Jun 8, 2026Zhiqiang Wu, Yitong Dong, Xian WeiDiffusion Models for Image Restoration Image Super-Resolution
1East China Normal University · 2Zhejiang University
Enhanced Self-Supervised Multi-Image Super-Resolution for Camera Array Images
Apr 8, 2026Yating Chen, Feng Huang, Xianyu Wu +2Autoregressive Image Super-Resolution Image Super-Resolution
Department of Precision Instruments, Tsinghua University, Beijing, 100084, China. · School of Mechanical Engineering and Automation, Fuzhou University, Fuzhou, 350108, China.

Feb 27, 2026J/K move · Enter open · S save

Altitude-Adaptive Vision-Only Geo-Localization for UAVs in GPS-Denied Environments

Xingyu Shao, Mengfan He, Liangzheng Sun, Chunyu Li+1

Department of Precision Instrument, Tsinghua University, Beijing 100084, China · School of Instrumentation Science and Optoelectronics Engineering, Beijing Information Science and Technology University, Beijing 100192, China · School of Aerospace Engineering, Beijing Institute of Technology, Beijing 100081, China

Cross-View Geo-Localization Map-Based Localization Visual Place Recognition GNSS-Denied Localization

Open paper AskPDF arXiv

Matching downward-looking unmanned aerial vehicle (UAV) images to georeferenced satellite or aerial map tiles supports local earth observation, map-based interpretation, and coarse geo-initialization when direct positioning signals are degraded or unavailable. A major difficulty in this cross-platform matching task is the scale mismatch caused by large variations in UAV altitude. To address this problem, we propose an altitude-adaptive aerial visual place recognition framework. The method first estimates relative altitude from a single downward-looking image by transforming the input into the frequency domain and formulating altitude estimation as a regression-as-classification problem. The estimated altitude is then used to crop the query image to a canonical scale, after which a classification-then-retrieval visual place recognition module performs coarse map localization. To improve retrieval robustness under varying image quality, we further introduce a quality-adaptive margin classifier and refine the final location by weighted coordinate estimation over the top retrieved candidates. Experiments on two synthetic datasets and two real-flight datasets show that the relative altitude estimation module improves downstream retrieval performance under substantial altitude changes. With our visual place recognition module, altitude adaptation improves average R@1 and R@5 by 41.50 and 56.83 percentage points, respectively, compared with using the same retrieval pipeline without altitude normalization, and the main retrieval pipeline runs at 13.3 frames/s on the workstation used for evaluation. These results indicate that relative altitude estimation provides a useful scale prior for cross-altitude UAV-to-map image matching using georeferenced remote sensing imagery.

Abstract

Image Super-Resolution

Explore similar work

Altitude-Adaptive Vision-Only Geo-Localization for UAVs in GPS-Denied Environments

TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution

Enhanced Self-Supervised Multi-Image Super-Resolution for Camera Array Images