A Novel Method Based on Deep Learning, GIS and Geomatics Software for Building a 3D City Model from VHR Satellite Stereo Imagery

The aim of the paper is to identify a suitable method for the construction of a 3D city model from stereo satellite imagery. In order to reach this goal, it is necessary to build a workflow consisting of three main steps: (1) Increasing the geometric resolution of the color images through the use of pan-sharpening techniques, (2) identification of the buildings’ footprint through deep-learning techniques and, finally, (3) building an algorithm in GIS (Geographic Information System) for the extraction of the elevation of buildings. The developed method was applied to stereo imagery acquired by WorldView-2 (WV-2), a commercial Earth-observation satellite. The comparison of the different pan-sharpening techniques showed that the Gram–Schmidt method provided better-quality color images than the other techniques examined; this result was deduced from both the visual analysis of the orthophotos and the analysis of quality indices (RMSE, RASE and ERGAS). Subsequently, a deep-learning technique was applied for pan sharpening an image in order to extract the footprint of buildings. Performance indices (precision, recall, overall accuracy and the F1 measure) showed an elevated accuracy in automatic recognition of the buildings. Finally, starting from the Digital Surface Model (DSM) generated by satellite imagery, an algorithm built in the GIS environment allowed the extraction of the building height from the elevation model. In this way, it was possible to build a 3D city model where the buildings are represented as prismatic solids with flat roofs, in a fast and precise way.

Impact of Base-to-Height Ratio on Canopy Height Estimation Accuracy of Hemiboreal Forest Tree Species by Using Satellite and Airborne Stereo Imagery

The present study assessed the large-format airborne (UltraCam) and satellite (GeoEye1 and Pleiades1B) image-based digital surface model (DSM) performance for canopy height estimation in predominantly mature, closed-canopy Latvian hemiboreal forestland. The research performed the direct comparison of calculated image-based DSM models with canopy peaks heights extracted from reference LiDAR data. The study confirmed the tendency for canopy height underestimation for all satellite-based models. The obtained accuracy of the canopy height estimation GeoEye1-based models varied as follows: for a pine (−1.49 median error, 1.52 m normalised median absolute deviation (NMAD)), spruce (−0.94 median, 1.97 m NMAD), birch (−0.26 median, 1.96 m NMAD), and black alder (−0.31 median, 1.52 m NMAD). The canopy detection rates (completeness) using GeoEye1 stereo imagery varied from 98% (pine) to >99% for spruce and deciduous tree species. This research has shown that determining the optimum base-to-height (B/H) ratio is critical for canopy height estimation efficiency and completeness using image-based DSMs. This study found that stereo imagery with a B/H ratio range of 0.2–0.3 (or convergence angle range 10–15°) is optimal for image-based DSMs in closed-canopy hemiboreal forest areas.

EXTRACTION OF CLOUD HEIGHTS FROM SENTINEL-2 MULTISPECTRAL IMAGES

Investigation of the focal plane assembly of the Sentinel-2 satellites show slight delays in the acquisition time of different bands on different CCD lines of about 0.5 to 1 second. This effect was already exploited in the detection of moving objects in very high resolution imagery as from WorldView-2 or -3 and also already for Sentinel-2 imagery. In our study we use the four 10-m-bands 2, 3, 4 and 8 (blue, green, red and near infrared) of Sentinel-2. In the level 1C processing each spectral band gets orthorectified separately on the same digital elevation model. So on the one hand moving objects on the ground experience a shift between the spectral bands. On the other hand objects not on the ground also show a slight shift between the spectral bands depending on the height of the object above ground. In this work we use this second effect. Analysis of cloudy Sentinel-2 scenes show small shifts of only one to two pixels depending on the height of the clouds above ground. So a new method based on algorithms for deriving dense digital elevation models from stereo imagery was developed to derive the cloud heights in Sentinel-2 images from the parallax from the 10-m-bands. After detailed description of the developed method it is applied to different cloudy Sentinel-2 images and the results are cross-checked using the shadows of the clouds together with the position of the sun at acquisition time.

Mapping Sub-Metre 3D Land-Sea Coral Reefscapes Using Superspectral WorldView-3 Satellite Stereoimagery

Shallow coral reefs ensure a wide portfolio of ecosystem services, from fish provisioning to tourism, that support more than 500 million people worldwide. The protection and sustainable management of these pivotal ecosystems require fine-scale but large-extent mapping of their 3D composition. The sub-metre spaceborne imagery can neatly produce such an expected product using multispectral stereo-imagery. We built the first 3D land-sea coral reefscape mapping using the 0.3 m superspectral WorldView-3 stereo-imagery. An array of 13 land use/land cover and sea use/sea cover habitats were classified using sea-, ground- and air-truth data. The satellite-derived topography and bathymetry reached vertical accuracies of 1.11 and 0.89 m, respectively. The value added of the eight mid-infrared (MIR) channels specific to the WorldView-3 was quantified using the classification overall accuracy (OA). With no topobathymetry, the best combination included the eight-band optical (visible + near-infrared) and the MIR8, which boosted the basic blue-green-red OA by 9.58%. The classes that most benefited from this MIR information were the land use “roof” and land cover “soil” classes. The addition of the satellite-derived topobathymetry to the optical+MIR1 produced the best full combination, increasing the basic OA by 9.73%, and reinforcing the “roof” and “soil” distinction.

Geodetic mass balance of Abramov Glacier from 1975 to 2015

Multi-decadal mass loss estimates are available for few glaciers of Central Asia. On Abramov Glacier (Pamir-Alay, Kyrgyzstan), comprehensive long-term glaciological measurements have been carried out from 1968 to 1999 and re-initiated in 2011. A climatological interpretation of this benchmark glacier in Central Asia requires bridging the gap between historical and renewed measurements. This is achieved here by computing the geodetic mass balance from 1975 to 2015 using previously unreleased Soviet aerial imagery and Pléaides stereo-imagery. During 1975–2015, Abramov Glacier lost 2.2 km2 (8.2%) of its area. The mean annual thickness change was − 0.43 ± 0.14 m a−1 for the period 1975–2015, corresponding to a volume change of − 0.45 ± 0.15 km3. The average specific geodetic mass balance amounts to − 0.38 ± 0.12 m w.e. a−1. The 1975–2015 glacier mass loss lies within the range of glaciological and geodetic mass-balance estimates that were previously published for disparate and shorter time intervals since 1968. This study covers a much longer time period than earlier geodetic estimates and demonstrates the capacity to geodetically constrain glacier change at high spatial resolution in Central Asia using historic aerial imagery and Structure from Motion techniques. Therefore, it could serve as a benchmark for future studies of regional mass change.