
Annual 64-dimensional AlphaEarth embeddings are available worldwide at 10 m. However, how far such a general representation transfers to a street-level quantity in parts of the city withheld from model training and used only for testing, and whether the embeddings improve on the products they replace remain untested. The embeddings are benchmarked on greenery segmented from 16 156 Mapillary views over 7531 50-m site-years in Istanbul, 2017–2024. Two further segmentation checkpoints agree on greenery at R²corr = 0.96 but on the road at only 0.38–0.64, restricting the primary claim to greenery. The protocol was frozen, and 119 of 551 2×2-km blocks were locked away before any model was fitted. On 2479 site-years in 75 withheld blocks, ridge regression on the 64 bands reaches 6.70 percentage points (pp) of mean absolute error compared with 8.92 for a naive mean (–2.23 pp; 95% interval, –3.04 to –1.39) and beats a strong observation-process baseline (–1.44 pp; –2.24 to –0.66). It does not beat a conventional reference of six Sentinel-2 bands with the normalized difference vegetation and built-up indices at matched points, supports, and learner on a season-matched composite (+0.006 pp; –0.44 to +0.43), and its predictive R² is lower (0.34 compared with 0.43). A boosted combination beats boosted AlphaEarth in both splits (–0.23 pp; 95% confidence interval, –0.34 to –0.13), reaching 6.21 pp and R² = 0.50. Block-shuffled embeddings collapse to the naive mean but neighboring-year embeddings do not, so the model reads place, not year; no temporal-change claim is supportable. Under a random split, a metadata-only model rises from R² = 0.03 to 0.32, the optimism of a nonspatial design. On a target that is this visible from orbit and in the one city tested, learned embeddings complement rather than replace conventional Earth observation. Code and data: https://github.com/yunusserhat/alphaearth_asprs.
Multi-temporal airborne lidar offers a potentially powerful tool for detecting forest mortality across scales. Lidar data are increasingly available through drone-based platforms and efforts like the U.S. Geological Survey 3D Elevation Program (3DEP). However, data sets may not always meet established best practices for lidar biomass estimation, including adequate penetration to the ground and collection during peak growing season. Further, variation in point cloud characteristics may complicate biomass change detection using multi-temporal data sets from different platforms. In this study, we assess using drone-based and 3DEP lidar for quantifying biomass and biomass change using data collected during two censuses of the 20-ha Paint Rock Forest Dynamics Plot. We explored using 3DEP elevation data to overcome limited ground information in drone-based lidar and using canopy height–based approaches to minimize differences from collection characteristics. We also considered the effect of quadrat size and plot biomass estimation method (stem-localized or crown-distributed) on correspondence between plot- and lidar-derived metrics. We found that both drone-based and 3DEP lidar estimated standing biomass with comparable accuracy to more optimal lidar data in previous studies. Lidar-predicted net biomass change was associated with plot-estimated values (Pearson r up to 0.57). However, compared to results for standing biomass, root mean square error for biomass change was considerably larger when expressed as a percentage of the mean value, highlighting important limitations of using multi-temporal lidar to characterize subtle biomass changes at fine resolutions and temporal scales. Stronger correlations were found when isolating mortality biomass loss of fallen canopy trees (r up to 0.70) than standing dead trees (r up to 0.29). Relationships for standing biomass and biomass change were consistently stronger for crown-distributed data, while relationships for mortality were similar for both biomass approaches.
Under adverse illumination, matchability scores may distinguish valid key points from ambiguous candidates less reliably. Fixed decision thresholds can then retain unreliable correspondences or discard useful candidates, destabilizing geometric estimation with random sample consensus. IPN-LightGlue is introduced to address this problem by coupling an illumination perception network (IPN) with adaptive matchability modulation. The IPN uses a frozen MobileNetV3-Small backbone to predict a normalized image-level illumination score, which is converted into a non-positive bias on LightGlue’s matchability logits. This design leaves the feature extractor unchanged and adds 74 000 trainable parameters. On the evaluation hardware, an isolated IPN forward pass takes 3.399 ± 0.829 ms per image, and the complete pipeline incurs a runtime overhead of 4.169% to 8.027% relative to LightGlue. Mean matching accuracy at a 3-pixel threshold (MMA@3px) increases from 79.9% to 85.2% on the HPatches illumination subset. For the multi-illumination data set, which isolates illumination-direction changes from viewpoint variation, MMA@3px rises from 85.2% to 89.5%, whereas mean corner error decreases by 16.4%. The method also reduces mean corner error from 59.75 pixels to 7.14 pixels on the HPatches viewpoint subset, despite the IPN being supervised only on radiometric variation. Neither contrast-limited adaptive histogram equalization nor zero-reference deep curve estimation, applied before matching, improves MMA@3px over LightGlue in these evaluations. The findings support illumination-aware matchability modulation as a way to improve correspondence accuracy and geometric stability without modifying the input image.
To address the limited field of view (FoV) of digital cameras in close-range photogrammetry, this study develops a novel, ground-based large-scale rotating photogrammetric system by integrating engineering surveying principles with photogrammetric theory. Results demonstrate that the approximate central projection images generated by this system can expand the coverage of a single camera station by more than five times while maintaining a root-mean square image stitching error less than 0.02 mm. The wide-format images generated by this system exhibit superior application accuracy and reconstruction efficiency in three-dimensional modelling compared with traditional small FoV images. The system reduces the root-mean-square error of control points and check points by 0.165 mm and 0.424 mm, respectively, while also significantly shortening the total reconstruction time. The proposed rotating photography system effectively mitigates the limited FoV of digital cameras and provides a novel technical solution for large-scale ground close-range photogrammetry. The systematic methodology for rotary image acquisition established in this work lays a theoretical and technical foundation for the practical engineering application of such systems.