An Altitude-Conditioned Spectral–Spatial Transformer for Reliable Cross-Altitude Crop Classification from Unmanned Aerial Vehicle Hyperspectral Imagery

Crop maps from unmanned aerial vehicle hyperspectral imagery increasingly guide site-specific field decisions, yet the classifiers behind them are almost always trained and evaluated at a single flight altitude. Because operational height varies with weather, battery and airspace limits, the ground sampling distance changes and map accuracy silently degrades at unseen heights, an error rarely measured before maps are trusted. We quantify this cross-altitude reliability gap and identify which design choices contain. Using a public open-field archive of three vegetable crops (tomato, cabbage and eggplant) imaged at six nominal heights (20–100 m), we build a reproducible protocol that trains at 30 m and evaluates at five withheld heights; because trustworthy labels exist only at 30 m, evaluation uses spectral-angle pseudo-labels together with chance-corrected metrics and an independent manual-validation set. We propose an altitude-conditioned spectral–spatial transformer with resolution augmentation and self-supervised pretraining, and compare it against convolutional, transformer and domain-adversarial baselines over five seeds. The proposed model gives the most reliable transfer, reaching a Cohen's kappa of 0.21 against 0.11–0.14 for the baselines, the best result on every chance-corrected metric and across all five seeds, with the firm evidence resting on the 25–50 m band. This advantage is relative and chance-corrected rather than absolute. Because the archive carries no ground truth, evaluation is referenced to Spectral Angle Mapper (SAM) pseudo-labels, so the species-level model ranking is SAM-consistent and awaits field-plot confirmation. An adversarial invariance component leaves flight height fully decodable, so informing the model of the altitude, not enforcing invariance, drives the gain, and no model exceeds a majority-class predictor in overall accuracy. Altitude-robust monitoring cannot yet be assumed: surveys should fly at consistent, documented heights and report chance-corrected accuracy. We release the protocol, pseudo-labeling pipeline, manual-validation set and all code.

Authors

Institutions

Publication Details

Journal
Black Sea Journal of Engineering and Science
Published
2026-11-15
DOI
https://doi.org/10.34248/bsengineering.2014931
Primary Topic
Remote-Sensing Image Classification
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An Altitude-Conditioned Spectral–Spatial Transformer for Reliable Cross-Altitude Crop Classification from Unmanned Aerial Vehicle Hyperspectral Imagery

Banu Ulu
Black Sea Journal of Engineering and Science
Remote-Sensing Image Classification
article

An Altitude-Conditioned Spectral–Spatial Transformer for Reliable Cross-Altitude Crop Classification from Unmanned Aerial Vehicle Hyperspectral Imagery

Banu Ulu
article en

Abstract

Crop maps from unmanned aerial vehicle hyperspectral imagery increasingly guide site-specific field decisions, yet the classifiers behind them are almost always trained and evaluated at a single flight altitude. Because operational height varies with weather, battery and airspace limits, the ground sampling distance changes and map accuracy silently degrades at unseen heights, an error rarely measured before maps are trusted. We quantify this cross-altitude reliability gap and identify which design choices contain. Using a public open-field archive of three vegetable crops (tomato, cabbage and eggplant) imaged at six nominal heights (20–100 m), we build a reproducible protocol that trains at 30 m and evaluates at five withheld heights; because trustworthy labels exist only at 30 m, evaluation uses spectral-angle pseudo-labels together with chance-corrected metrics and an independent manual-validation set. We propose an altitude-conditioned spectral–spatial transformer with resolution augmentation and self-supervised pretraining, and compare it against convolutional, transformer and domain-adversarial baselines over five seeds. The proposed model gives the most reliable transfer, reaching a Cohen's kappa of 0.21 against 0.11–0.14 for the baselines, the best result on every chance-corrected metric and across all five seeds, with the firm evidence resting on the 25–50 m band. This advantage is relative and chance-corrected rather than absolute. Because the archive carries no ground truth, evaluation is referenced to Spectral Angle Mapper (SAM) pseudo-labels, so the species-level model ranking is SAM-consistent and awaits field-plot confirmation. An adversarial invariance component leaves flight height fully decodable, so informing the model of the altitude, not enforcing invariance, drives the gain, and no model exceeds a majority-class predictor in overall accuracy. Altitude-robust monitoring cannot yet be assumed: surveys should fly at consistent, documented heights and report chance-corrected accuracy. We release the protocol, pseudo-labeling pipeline, manual-validation set and all code.

Black Sea Journal of Engineering and ScienceVol. 9(6)
Kayseri Üniversitesi (TR)
Remote-Sensing Image Classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon