Planet4Stereo: A Photogrammetric Open-Source Pipeline for Generating Digital Elevation Models for Glacier Change Monitoring Using Low-Cost PlanetScope Satellite Data

Abstract

Monitoring volumetric glacier change requires cost-effective and accessible methods to generate multi-temporal digital elevation models (DEMs). We present Planet4Stereo, an open-source photogrammetry pipeline developed to generate DEMs from low-cost PlanetScope images, exploiting the high temporal repetition rate of the constellation for stereo reconstruction. Our approach enables multi-temporal 3D change detection using the freely available NASA Ames Stereo Pipeline (ASP), making the pipeline particularly valuable for geoscientists. We applied Planet4Stereo in two case studies: the Shisper glacier (Karakoram, Pakistan) for surge investigation and the Bøverbrean glacier (Smørstabb Massif, Norway) for change detection. The results from Shisper are in good agreement with previous studies using the same images but proprietary methods. The accuracy of the DEM of Bøverbrean was evaluated using high-precision LiDAR data, revealing varying deviations across terrain types, with higher errors in steep shadowed areas. Additionally, the change detection analysis confirmed the expected glacier retreat. Our results show that Planet4Stereo produces DEMs with comparable accuracy to commercial software and is freely accessible and easy to use. As both ASP and the PlanetScope satellites evolve, future work could refine the pipeline’s stereo-matching capabilities and evaluate its performance with next-generation satellite data.

1. Introduction

Due to environmental changes and human activities, there is an increasing need for continuous monitoring of Earth’s surface to observe phenomena such as urbanization, landslides, and glacier activities. The use of multi-temporal DEMs allows for the determination of, for example, changes in glacier extent and mass balances [1,2,3,4]. Freely available DEMs provide global elevation information but at a ground sampling distance (GSD) of approximately 30 m per pixel using merged data captured over several days, e.g., the SRTM DEM [5] or the Open Copernicus DEM [6]. Moreover, the temporal resolution of such DEMs is frequently inadequate for the analysis of dynamic processes. The well-known ASTER system, carried by the Terra satellite, captures panchromatic stereo images for 3D processing with revisit times of 4–16 days but with limited GSD of about 15 m per pixel. Commercial very high-resolution (VHR) satellite images offer stereo capabilities, e.g., WorldView-2/3 [7,8,9,10] or Pleiades [11], and enable DEMs with GSDs in the range of about a meter. However, the data are rather expensive, especially in terms of change detection, as at least two DEMs taken at two different times covering the time span to be studied are required. The same applies to on-demand constellations such as Pléiades Neo, which are capable of acquiring scheduled VHR stereo image pairs.

PlanetScope, operated by Planet Labs PBC, is a constellation of Earth observation satellites designed to continuously monitor Earth’s surface. The company’s goal is to provide frequent and easily accessible satellite imagery for various applications, such as environmental monitoring, agriculture, urban planning, and disaster prevention. Particularly interesting for many scientific questions is the fact that the data are free of charge for non-commercial purposes, e.g., via the ESA Earthnet Programme (https://earth.esa.int/eogateway/catalog/planetscope-full-archive, accessed on 18 February 2025) or the Planet Education and Research Programme (https://www.planet.com/industries/education-and-research/, accessed on 18 February 2025).

The PlanetScope constellation currently consists of approximately 130 low-cost cubesats, called Doves, which operate in sun-synchronous ascending and descending orbits while Earth rotates underneath, allowing for gap-less mapping of Earth’s surface at daily resolutions (see Figure 1).

Doves are composed of adapted components from miniaturized consumer electronics [13]. The high-resolution push-frame imaging system consists of a frame camera charge-coupled device (CCD) sensor and a mirror telescope lens capturing Earth’s surface with a GSD of 3–4 m. In this push-frame system, wavelengths are separated from each other by filter masks so that specific parts of the sensor are only affected by certain wavelengths, enabling multi-spectral imaging. The individual sub-frames are assembled by band co-registration into complete scenes according to the sensor resolution in the factory post-processing. Depending on the cubesat generation, the (full) CCD sensor has a spatial resolution of approximately 6660 × 2220 pixels (first Dove Classic generation) or 6660 × 4400 pixels (second Dove-R and third SuperDove generation), which results in footprints of 24 × 8 and 24 × 16 kilometers, respectively. The spectral resolution of Dove Classic and Dove-R was limited to RGB and near-infrared (NIR). SuperDove additionally resolves red edge, coastal blue, yellow, and a small band in the range of green. Since April 2022, the Dove Classic and Dove-R satellites have been decommissioned, but the data collected between June 2016 and April 2022 by Dove Classic and between March 2019 and April 2022 by Dove-R are still available as archived data. SuperDove scenes have been available since March 2020 and are still captured daily [14]. All scenes are taken with slightly off-nadir view angles (0–5°, depending on latitude) at altitudes ranging between 475 and 525 km. The overlap along the track is about 10%, with a time interval between consecutive images of about one second. To ensure void-less coverage, the swaths of neighboring cubesats overlap by several kilometers in cross direction (approximately 40%; time interval of approximately 90 s, e.g., [15]).

The PlanetScope scene (PSS) images are available as basis scene products (Level 1B) and orthorectified scene products (Level 3B). Level 1B products provide scaled top-of-atmosphere radiance values with radiometric and geometric sensor corrections applied. According to the Planets product description, the latter imply sensor optic corrections (lens distortion) and co-registration of the captured strips. Level 1B scenes are not orthorectified but accompanied by rational polynomial coefficient (RPC) describing the imaging geometry by a rational functional model (RFM), which is particularly important for 3D reconstruction when the intrinsic (physical) rigorous sensor model is unknown. The RPC consist of four polynomials, each containing 20 coefficients, used to transform geodetic coordinates (longitude, latitude, and height above the datum) into image pixel coordinates.

The reader is therefore advised to select this processing stage when obtaining data if s/he wishes to use the approach presented below. More details on the PlanetScope mission, constellation, and data products are given in [12,14,15,16].

Stereo Limitation

Due to the temporal convergence of multi-temporal images taken on staggered orbits with slight off-nadir angles, the PlanetScope constellation represents an interesting alternative to known commercial missions for generating 3D surface models with very high temporal and fairly high spatial resolutions to study, for example, glacier dynamics, e.g., [15]. However, the mission is not designed for stereo evaluation by default. This option results purely from the extremely high temporal repetition rate of the cubesats, which capture the same area from slightly different perspectives several times a day/week. The limitations in the stereo capability result from rather short baselines between potential image pairs due to small off-nadir angles, small footprints, and large across-track overlaps. Ref. [17] reported base-to-height ratios in the range of 1:10 to 1:100, where photogrammetry strives at ratios in the order of 1:1. A small baseline causes narrow intersection angles of two visual lines representing image features and intersecting in an object point, thus reducing the depth information precision. On the other hand, stereo image matching becomes much easier due to similar perspectives and fewer occlusions, especially in mountainous areas, e.g., [18]. Both allow for a certain amount of compensation, which can enable the creation of a reliable DEM using highly redundant image information from multi-temporal images, even if the stereo configuration is not optimal.

In the following, we provide a literature review of studies that have utilized PlanetScope data for DEM generation. Subsequently, we present our methodology, detailing the Planet4Stereo workflow based on the ASP. In the Experiments Section, we apply Planet4Stereo to generate 3D surface models for both large- and small-scale areas. The accuracy potential of the pipeline is evaluated by comparing the generated models with reference DEMs.

2. Related Work

The potential of using PSS images for 3D stereo photogrammetry and 4D change detection was already shown in a few studies, e.g., Refs. [16,17,19,20], with similar workflows and results.

The first study, to our knowledge, focusing on DEM generation from PlanetScope data for geomorphological analyses was published in [17]. The proposed method implies a stabilization of the image block applying bundle-block adjustment, here called bias correction. This optimizes the RFMs of the individual images and is followed by pairwise stereo reconstruction applying semi-global matching (SGM) in object space. The resulting point clouds are rasterized into a single DEM with a GSD of 5 m/px using the median altitude in overlapping areas. The DEM is finally co-registered to an already existing DEM to correct systematic shifts commonly resulting from bundle-block adjustment without using ground control points (GCPs) that impact the georeferencing. Details on the implementation, tools, or used libraries were not provided. The method was applied to generate DEMs in high mountain and partial glaciated areas, e.g., Volcano Teide (steep terrain and sparse vegetation; reference DEM from LiDAR) and Naga Parabat Massif (glaciated high mountains and no vegetation; ALOS World DEM (AW3D30) as reference DEM).

The authors of Ref. [16] selected suitable image pairs in a first step by examining all the available images of a region of interest (ROI) captured over a certain time period, selecting image pairs with maximum overlap and the best convergence angles. They performed bias correction followed by point cloud generation using Agisoft Metashape’s Dense Matching function. The generated point cloud is resampled to a DEM with a GSD of 9 m. The proposed method for bias correction was implemented in the software COSI-Corr, which was originally published as a plug-in for the commercial remote sensing software ENVI from NV5 Geospatial Solutions [21]. Since 2022, COSI-Corr has also been available as a standalone version (e.g., [22]). The approach was used to detect changes in glacier topography before and after a surge event. In a more recent study, Ref. [23] provided more details on how to select suitable image pairs for this operation: The minimum number of overlapping images should be greater than five, the base-to-height ratio should be better than 1:7 with an overlap of at least 60%, and the images should be captured from opposite satellite azimuths, i.e., ascending and descending orbits as well as in short time spans to reduce, for example, weather and seasonal influences on the image representations.

In the work of [20], almost all the available PSS images captured in a specific time span were used to generate a DEM of an area with various topographic highlights, e.g., cities, forested hills, and steep mountains. They also performed bias correction as a first step via bundle-block adjustment followed by pairwise stereo reconstruction applying SGM. The individual point clouds were rasterized to DEMs and analyzed for height errors in correlation with the original convergence angle of the stereo pair. Those DEMs generated from image pairs with convergence angles larger than 6° showed fewer errors and were merged to a single DEM using the median height. Details on the algorithms, implementation, or used tools were not provided. The authors of Ref. [19] evaluated the potential of PlanetScope images for 3D applications with a specific focus on 4D change detection in urban, rural, forested, and glacial areas. They applied an in-house-developed multi-view stereo (MVS) pipeline based on [24] and compared the results with reference 3D data from LiDAR. Similar to [20], they found that accurate DEMs can be generated using image pairs with convergence angles larger than 8°.

The accuracy investigations of [17] show normalized median of absolute deviations (NMADs) of the DEMs derived from PlanetScope of 4.1 m compared to a LiDAR DEM and 3.9 m compared to the AW3D30 over stable terrain. The authors of Ref. [16] compared the calculated PlanetScope DEMs to models generated by stereo photogrammetry using commercial stereoscopic VHR satellite images, i.e., GeoEye-1 and WorldView-2, as well as to the SRTM DEM in stable areas. The NMADs are about 7 m and 12 m compared to the VHR DEM and the SRTM DEM, respectively. The larger differences to the SRTM DEM may result from its lower resolution and overall accuracy [25]. No comparison to independent reference data was conducted by [23], but a relative error of 10 m, calculated by comparing the PlanetScope DEMs with each other in stable terrain, was reported. Mass movements of the surging glacier could be clearly identified in both studies. Ref. [20] reported a root mean square error (RMSE) of 5.5 m comparing the PlanetScope DEM to LiDAR data. Similarly, the authors of Ref. [19] reported an RMSE of about 5 m, also comparing the derived PlanetScope DEM to LiDAR data. They found that volumetric changes can be detected when the signal exceeds 8–10 m in a vertical direction. All the studies emphasize that the DEM generation using PSS images strongly depends on the data availability regarding suitable base-to-height ratios, which in most cases is very small due to the image configurations. However, this can be compensated to some extent due to the (often) abundance of possible image pairs with similar footprints from slightly different perspectives and thus good matching and beam intersection properties. The time span in which possible image pairs are taken should always be chosen in relation to the dynamics of the ROI to reduce errors induced by changing objects. The studies also found that the supplied RFMs describe the imaging geometry very well, and only minor biases in the range of 0.05–0.2 px are to be expected, which could be solved with the help of bundle-block adjustment, e.g., [23,26].

In a nutshell, DEMs from multi-view PlanetScope imagery show height accuracy potentials of about 1–2 px, i.e., 4–8 m. They can provide valuable information for studying volumetric changes at the global scale and are well suited to investigate Earth dynamics like landslides and glacier changes in four dimensions. All the workflows described in the literature are similar but require either non-published in-house implementations or commercial software solutions for the processing of the data. In this paper, we present an easy-to-use open-source stereo processing pipeline for generating DEMs from multi-temporal PlanetScope images, which requires only a few steering parameters and is aimed at a target group from applied geosciences and glaciology in particular.

A very promising open-source toolbox for photogrammetric 3D reconstruction using space-borne data is ASP. ASP includes optimized algorithms for image block pre-processing (e.g. bundle-block adjustment), stereo matching (e.g. SGM), and post-processing (e.g. error assessment, outlier removal, point cloud registration, and DEM rasterization). All the tools can be loosely combined and parameterized for a wide variety of input data and application purposes. Details on ASP are given in [27,28]. ASP has already been successfully deployed for processing data of SkySat imagery, a satellite constellation also launched by Planet Labs in 2013. Moreover, [29] published the open-source pipeline SkysatStereo using tools from ASP for DEM generation from SkySat-C image triplets and video sequences (GSD of the panchromatic images approximately 0.7 m). The pipeline was successfully applied to generate DEMs of the Mount Rainier glaciers, with GSDs of 2 m showing vertical accuracies of 2–3 m. Recently, the authors of Ref. [12] presented an approach in which ASP was used to generate a DEM from PlanetScope data. They referred to a specific use case and used only two images with a rarely achievable configuration and large convergence angle to generate a relatively small and coarse-resolution DEM with a GSD of 30 m and demonstrated the suitability of PlanetScope data and ASP for stereo reconstruction.

3. Methods

Motivated by SkysatStereo, we introduce Planet4Stereo—an automated pipeline using ASP for the stereo photogrammetric processing of multi-temporal PlanetScope images. Planet4Stereo is implemented in Python 3.10, merely requiring the installation of ASP version 3.4.0 and a few Python libraries commonly known for geo-processing, like Geopandas, GDAL, or Rasterio. It was tested with Ubuntu 22-04 LTS running in the Windows Subsystem for Linux (WSL).

The pipeline is illustrated in Figure 2 following a similar scheme as described in the literature (see Section 2). As an initial step (not illustrated in the figure), the operator should download the requisite input data, i.e., the PSS, a coarse-resolution DEM (hereinafter referred to as ‘approximation DEM’), and, if required, glacier mask data.

This approximation DEM, as outlined in the ASP documentation, facilitates the calculation of the disparity, particularly in topographies characterized by significant inclines [27]. One option is the use of the Copernicus World DEM (GLO-30) with a GSD of 30 m [6]. The pipeline commences with the selection of the potential stereo pairs, whereby the minimum necessary overlap and convergence angle are analyzed. The resulting image block configuration is optimized through the utilization of bundle-block adjustment for bias correction. Even if the images already have similar perspectives, they are orthorectified using the approximation DEM to further improve image correlation in the stereo matching process. In this so-called ortho workflow, the approximation DEM serves as an initial guess when calculating the disparities and subsequently the 3D point cloud. This workflow is highly recommended by ASP with steep terrain as it occurs in glaciological applications. The pairwise-calculated 3D point clouds are aligned to the given approximation DEM to correct for any shifts in georeferencing that may have occurred due to bundle-block adjustment without using GCPs. This is followed by rasterizing all pairwise-calculated point clouds to DEMs. In this step, the ASP option is used to create orthorectified images in the area of the respectively generated DEM. In a later step, all images orthorectified in this way can be used to create an orthomosaic of the entire scene. Additionally, an error map is generated to represent the mean intersection error of the reconstructed 3D surface points within each grid cell. This map facilitates the investigation of potential systematic errors. The pairwise-generated DEMs, error maps, and orthophotos are then combined into a mosaic to produce the final products: a DEM mosaic, a merged error map, and an orthophoto mosaic covering the ROI.

The following sections provide a detailed overview of the methodology, with a particular emphasis on the tools utilized within the ASP environment. Tools are denoted using (brackets and italics).

3.1. Input Data

Depending on the ROI and the expected dynamics that could constrain the selection of multi-temporal image pairs, all available images within the specified time period should be reviewed, for example using the PlanetExplorer web interface. It is advisable to prioritize scenes with a finalized or standard processing state. Additionally, cloud coverage should be minimal, ideally ≤10%. The data must be downloaded as Level 1B products, which include metadata in JSON format required for image-pair selection and RPCs necessary for reconstructing the image geometry during subsequent bundle-block adjustment and stereo processing. The approximation DEM is required to be in geographic coordinates with ellipsoidal heights above the WGS84 reference ellipsoid. Highly variable areas, such as glaciers, should be excluded during the re-alignment of the generated 3D point clouds. This can be achieved by utilizing vectorized mask data, such as those provided by RGI [30].

3.2. User Settings

With respect to the target user domain, the necessary settings were consciously limited to ensure straightforward usability. The user only needs to specify the paths to an empty working directory and to the datasets mentioned, i.e., to the PSS directory, storing images, metadata, and RPC files, to the approximation DEM and to the glacier masks, provided as ESRI shape file. If the approximation DEM uses orthometric heights, Planet4Stereo with ASP supports direct conversion to ellipsoidal heights by specifying the geoid model as an optional argument, e.g., EGM96 for SRTM or EGM2008 for GLO-30, eliminating the need for prior transformation.

3.3. Expert Settings

The default settings provided by Planet4Stereo and ASP have been validated through multiple successful glacial 3D reconstructions using the pipeline (see Section 4). These settings require no modifications. The chosen default values for certain parameters were determined empirically: the subpixel kernel size of 35 strikes a balance between achieving high subpixel accuracy and maintaining manageable computational costs, particularly in low-texture or low-contrast regions. Similarly, the correlation kernel size of 7 has proven optimal for capturing sufficient local context to resolve ambiguities during the matching process without overly smoothing fine details. These defaults have consistently yielded robust results in both mountainous and glaciated terrains. However, experienced users may adjust the parameters listed in

Table 1. Expert settings.

Parameter	Type	Description	Default
pss_band	int	Specifies the band used for stereo reconstruction. The Near-Infrared (NIR, Band 4) is recommended for improved contrast in saturated areas, such as snow.	4
no_ortho	bool	Disable orthorectification before stereo reconstruction (not recommended).	False
elevation_tolerance	float	Elevation tolerance [m] for filtering out coarse outliers in bundle-block adjustment	500.0
min_convergence_angle	float	Minimum convergence angle [deg] between two PSS images.	4.0
min_overlap_percent	float	Minimum overlap percentage between two PSS images (0.0–1.0 ≡ 0–100%).	0.1
subpx_kernel	int	Subpixel kernel size ([px], use larger values for Bayes EM or low-texture images).	35
corr_kernel	int	Correlation kernel size ([px], use odd value, 3–9 for SGM or MGM methods).	7

to further optimize individual results. Additional fixed settings can be reviewed directly within the pipeline script. In addition, several debugging options are available, allowing the deactivation of specific processing steps for code troubleshooting. A detailed discussion of these functions is beyond the scope of this paper. Readers interested in exploring these features are encouraged to check the Python code and accompanying documentation.

3.4. Pre-Processing

Prior to the initiation of the actual pre-processing phase, all input data are transferred to a designated working directory. If required, the approximation DEM is converted to ellipsoidal heights and masked with the use of, for instance, the outlines derived from the RGI for later point cloud co-registration. The selection of stereo image pairs is conducted by parsing the metadata describing the footprints of the PSS images and their corresponding RPCs. This allows for the pairwise calculation of the image overlap and the convergence angle between all possible scene combinations, as described by [29,31].

3.5. Bundle-Block Adjustment

As indicated in the literature, the PSS images are well oriented and described by RFMs, represented through RPCs. The RPC bias correction, i.e., correction of errors in camera position and orientation, could be solved via bundle-block adjustment (bundle_adjust) and is expected to result in a residual bias of less than one pixel. The adjustment minimizes the reprojection error between estimated back-projected pixel positions of 3D object points and their actual measured positions in the images [28]. For this purpose, feature point detection and matching are performed for all selected image pairs using the default OBAloG algorithm [32]. These matched points are triangulated into 3D surface points using the provided RPCs. To optimize performance, the number of interest points is limited to the top 5000. Given the anticipated minimal bias, the weighting is configured to allow only minor adjustments to the sensor poses. Outliers in the image matching are excluded through RANSAC, which is used in the course of computing the fundamental matrices. For triangulation, outliers are filtered by rejecting 3D points with intersection angles below the specified minimum convergence angle. Additionally, 3D points are classified as outliers if their calculated elevations exceed the possible range, which is determined by the minimum and maximum elevations of the approximation DEM ± an elevation limit threshold (see Table 1). Images that could not be aligned during bundle-block adjustment are rejected from further processing. Consequently, potential corresponding image pairs that include these images are also excluded. The default settings have been retained for all other parameters used in (bundle_adjust). The estimated correction values for the sensor poses are output as translation and rotation adjustments for each scene. These corrections are automatically applied by ASP to the respective RPCs during subsequent processing steps.

3.6. Stereo Processing

By default, all images of the aligned image block are orthorectified (mapproject) before stereo reconstruction using the optimized RFMs and the approximation DEM. Orthorectifying the PSS images in advance results in more similar perspectives, which likely improves both the number and quality of stereo matches, even at the cost of slightly higher computing times during the initial image matching step, running parallel_stereo. The disparity remains as residual information because the DEM does not perfectly model the actual surface (e.g., due to its coarser resolution, different acquisition times, or inherent inaccuracies). ASP uses these residual disparities, together with the optimized camera models, to reconstruct the actual 3D surface by updating the approximation DEM. In the absence of orthorectified images and approximation DEM (e.g., when using the Expert setting no_ortho), image rectification is carried out using an affine epipolar transformation. This step is mandatory for subsequent image matching that requires rectified images. The information from the camera models is then used to perform traditional triangulation, transferring the disparity information into object space. Disparity is measured using an adaptation of the SGM algorithm, called more global matching (MGM) [33]. Compared to SGM, MGM offers significant advantages in low-texture scenarios: by aggregating matching costs over a larger spatial context, striping artefacts and false matches in low-texture areas are significantly reduced. This improves the robustness of the disparity computation despite higher computational cost. Subpixel refinement is applied using Bayes Expectation-Maximization (Bayes EM), as recommended by ASP, which provides the highest-quality results. The kernel sizes for correlation and subpixel refinement were selected based on empirical evidence and align with the default values. These values are optimal for mountainous and glaciated regions. In areas with limited depth variation and minimal contrast (i.e., texture) in the PSS, it is recommended to slightly increase the kernel sizes in the Expert Settings to reduce ambiguities in the correlation. However, excessive kernel sizes may lead to a loss of sharp edges in the resulting disparity map. The pairwise-generated 3D point clouds undergo a filtering process, where isolated blobs with fewer than nine pixels are identified and removed.

3.7. Point Cloud Alignment

Running bundle-block adjustment without GCPs can lead to systematic shifts, which affect the georeferencing of the image block. To mitigate potential discrepancies, the calculated 3D point clouds are individually co-registered to the approximation DEM in stable areas using point-to-plane alignment (pc_align). The maximum allowed displacement is set to 2000 m. While this may seem like a large threshold, it is necessary to accommodate correctly calculated point clouds that may include a few remaining outliers. These outliers can cause significant deviations from the approximation DEM, which the alignment process must still be able to handle effectively.

3.8. DEM Generation and Mosaicking

The aligned and co-registered 3D point clouds are individually rasterized using the (point2dem) function, thereby preserving the native resolution of the PSSs in geographic coordinates. This is possible due to the fact that a high number of points will fall into one cell of the raster as a result of the multi-view principle in the region of interest, yielding high redundancy. During this process, the intersection errors provided by ASP in meters for each 3D point are analyzed to detect and remove outliers. Points with intersection errors exceeding three times the 75th percentile of all points in the error map are identified as outliers and excluded from further processing. Furthermore, isolated points are removed.

Small gaps in the resulting DEMs are interpolated using a search radius factor of three pixels, and a 11x11 median filter is applied to remove spikes. Error maps are generated during this step, showing the intersection error for each point in the corresponding DEM. These maps depict the mean divergence of ray intersections for object points within each raster cell, defined as the shortest distance between two image rays that theoretically intersect at a single point in object space. They provide a quantitative measure of the uncertainty in 3D object point calculations and facilitate the detection of potential systematic errors. Moreover, the point2dem tool facilitates the orthorectification of, by default, the left image from each stereo pair, which is subsequently used to generate an orthophoto mosaic. This is not the core task of Planet4Stereo but may be an interesting by-product for the user. The DEMs, error maps, and orthophotos are finally merged using functions of the rasterio library for orthophoto mosaicking and ASP’s dem_mosaic for DEM and error map mosaicking.

3.9. Accuracy Assessment

To evaluate the quality and precision of the computed DEM, it is recommended to perform DEM of differences (DoD) analyses using, if available, independent 3D reference data in areas that are predominantly stable. This approach was commonly employed in related studies (see Section 2). It should be noted that this step is not part of the Planet4Stereo pipeline. To address horizontal and vertical shifts usually inherent in DEMs, we employ demcoreg, a library of tools developed by [28]. Removing relative offsets is essential for reliable 3D change detection and, equally, for the accurate evaluation of the calculated PlanetScope DEMs. Within demcoreg, we use a slightly adapted version of the dem_align tool, which implements the co-registration algorithm described by [4]. Glacier regions are excluded using outlines from RGI v7.0. For consistency, the reference DEM and the calculated PlanetScope DEM are resampled to the average resolution of both datasets. This is achieved through statistical analysis, using key metrics such as the median, mean, standard deviation (StdDev), and NMAD.

4. Practical Experiments

Planet4Stereo was tested for change detection in two glacier areas of varying scales and characteristics. The first case study focused on the Shisper glacier (Karakoram, Pakistan), selected for its susceptibility to surges and prior use by [16] to validate PlanetScope DEM processing methods. Using the same PlanetScope scenes as provided by the authors, we evaluated the accuracy of the DEM through DoD calculations against reference data. Our results were similar to [16] in a stable region near Ali Abad/Karimabad village. Additionally, we performed an independent change detection analysis and benchmarked our findings against their published results. The second case study focused on the Bøverbrean glacier in the Smørstabb Massif, Jotunheimen, Norway. Glaciers in this region have experienced significant retreat in recent years, prompting ongoing monitoring by Norwegian authorities (e.g., [34,35]). For our analysis, we utilized nearly cloud-free PlanetScope scenes from August and September 2021 to generate a DEM. We evaluated this model by comparing it to a LiDAR dataset captured in 2020. Additionally, we conducted change detection by comparing the PlanetScope DEM with a dataset generated from aerial images by aerotriangulation in 2013.

4.1. Shisper Glacier

Shisper glacier, situated in the Karakoram range of northern Pakistan, is known for its dynamic and fast-moving behavior, characterized by rapid changes and periodic surges. A notable surge in 2018 led to the formation of a large ice-dammed lake, posing a significant threat to the nearby Ali Abad village due to the risk of catastrophic flooding from a potential outburst. This event has drawn considerable scientific attention, with monitoring efforts combining satellite observations and on-ground inspections to provide early warnings to local communities. The glacier’s activity exemplifies the dynamic and unpredictable nature of Karakoram glaciers, influenced by climate change and geological processes, making it a key focus in glaciological research (e.g., [36,37,38,39]). Ref. [37] highlighted the use of multi-temporal DEMs, derived from PlanetScope data, to analyze the 2018 surge event, referencing the methodology established by [16]. In their work, Shisper glacier was mapped in 2017 and 2019 to validate the proposed method (see Section 2) and to quantify the surge event using DoD analysis. This makes Shisper glacier an ideal case study for comparing Planet4Stereo with an established commercial-software-based method as both utilize the same PlanetScope scenes and nearly identical regions for accuracy assessment. For details on the scene IDs and data acquisition parameters, refer to Tables A1 and A2 in [16] (open access). The study area and cloud-free PSS images used for 3D modeling in 2017 (16 scenes captured over 11 days) and 2019 (14 scenes captured over 12 days) are shown in Figure 3 and Figure 4, respectively.

It is important to note that the scenes selected for 2019 differ by two scenes compared to [16]. Specifically, we focused on the flight strips of satellites 1021 and 1006, including images captured at 05:22:32 and 05:24:23, respectively. These scenes provide better coverage of the ROI and the Ali Abad village, facilitating validation purposes. Conversely, we excluded the scenes taken at 05:22:29 and 05:24:20, used by the original authors, as they primarily cover the northern glacier area outside the ROI. In [16], a presumably stable area of approximately 25 km² around Ali Abad village (outlined in red in Figure 3 and Figure 4) was used to assess the accuracy of the generated DEMs. This same area was considered in our subsequent accuracy analysis.

4.1.1. Processing—Shisper Glacier

Planet4Stereo was run using the NIR band of the PSS considering image pairs with a minimum overlap of 10% and a convergence angle of at least 4°. The GLO-30 DEM served as the approximation, while glacier regions were masked using RGI outlines during the alignment of pairwise-calculated point clouds to mitigate potential shifts resulting from bundle block adjustment without the use of GCPs.

Table 2. Shisper glacier DEM processing settings.

	PlanetScope DEM 2017	PlanetScope DEM 2019
Configuration
Number of images	16 (NIR)	14 (NIR)
Minimum overlap (%)	10
Minimum convergence angle (°)	4
Ortho workflow	yes (GLO-30 DEM)
Bundle-block adjustment
Number of triangulated points (sparse)	7762	4160
Mean residuals of images (px)	0.3	0.3
Mean residuals of triangulated points (px)	0.3	0.3
Stereo reconstruction
Mean intersection error (m)	1.4 (max: 11.8)	1.6 (max: 13.7)
DEM sampling
DEM coverage (km2)	600.8	444.1

and Figure 5 present the resulting DEMs and corresponding statistics.

Mean residuals at the camera positions from bundle-block adjustment and for triangulated image points (detected in at least two images) were approximately 0.3 pixels for both years. These values align with the literature expectations and closely match the results reported by [16]. The intersection errors in the error maps are randomly distributed, with higher errors observed in snow-covered regions and areas with steep terrain. The average intersection errors are 1.4 m in 2017 and 1.6 m in 2019, corresponding to 0.3 pixels and 0.4 pixels, respectively, relative to the mean GSD of the PSS, which is 3.9 m. Note that the DEM generated in 2019 covers the ROI of Shisper glacier well (as in 2017) but with some gaps in the upper mountain regions. This is attributed to the fact that the image coverage focused on a better representation of the ROI, which was at the expense of the image coverage of the higher areas. No ascending scenes were usable covering the ROI.

4.1.2. Evaluation—Shisper Glacier

To assess the accuracy of the generated DEMs, DoD analyses were conducted in specific regions, comparing the PSS DEMs with the GLO-30 DEM used as a reference. These analyses focused on stable areas such as Ali Abad village, the broader Ali Abad region encompassing steep northern mountainous terrain, and a section of mountainous terrain partially covered by snow in the central study area. These selections aimed to investigate the impact of snow cover and steep terrain on DEM accuracy. Additionally, intersection errors were examined to assess triangulation precision. The GLO-30 DEM demonstrates global absolute and relative vertical accuracies of <4 m in steep terrain within an empirical 90% quantile [6]. However, an analysis according to individual region types shows systematically higher deviations in mountainous regions. This is particularly true for regions of the “Rock and Ice” type, into which the Muchowar and Shisper glacier areas are categorized. Here, the empirical 90% quantile is given as <14 m, which must be taken into account when interpreting the following DoDs [40].

The DoD analysis of the PSS DEM from 2017 in Ali Abad/Karimabad, as shown in Figure 6, revealed a mean deviation of 0.99 m (median of 0.43 m), a StdDev of 9.44 m, and an NMAD of 7.94 m in the anticipated stable region. These findings align closely with those of [16], who reported a mean deviation of 0.3 m, a StdDev of 11.0 m, and an NMAD of 12 m in comparisons to SRTM data for the same year. A slight negative systematic deviation along the Hunza River is noticeable, which will be discussed in more detail below.

Analysis of the Aliabad/Karimabad patch (see Figure 7) highlights significant DoD deviations in rock slope regions north of the settlements, likely due to low texture, shadowing, and steep ray intersection angles impacting 3D reconstruction accuracy. In addition, the lower quality of the reference DEM in this area must be considered. The error map, however, shows no anomalies in the northern slopes but reveals a systematic pattern in the south, where ray intersection errors are higher. This may result from southern areas being reconstructed using only two DEMs from stereo pairs captured with similar configurations (see Count, Figure 3 and Figure 7). The ray intersection angles in this area were likely less favorable than in the northern area. Regions with multiple paired DEMs benefit from greater redundancy, improving accuracy.

In the southern region near the Hunza River, a slight systematic negative deviation is observed, with the river lying several meters below the village, forming a steep edge in the DEM. The exact cause of this deviation remains unclear. Possible contributing factors include limited textures, as seen in the orthophoto, and suboptimal ray intersection angles. Terrain changes, such as landslides, between the GLO-30 DEM and the PSS 2017 DEM cannot be ruled out, nor can uncertainties in the GLO-30 DEM itself. High-resolution reference data collected concurrently with the PSS images could provide clarity.

The analysis of the Muchowar glacier patch (see Figure 8) reveals significant DoD deviations in steep terrain areas with pronounced shadowing or snow coverage in the PSS (see orthophoto), potentially linked to snow thickness variations. Snow-covered regions generally pose fewer issues for 3D reconstruction itself than expected, as indicated by the error map, apart from some higher ray intersection errors in the eastern section, typically at steep edges. This may be due to the use of the NIR band as recommended for icy/snow-covered terrain. Notably, along the Muchowar glacier tongue, systematic ray intersection errors are observed, likely due to glacier movement during the 11-day image acquisition period, causing point displacement and triangulation inaccuracies.

The 2019 DoD analysis of the PSS DEM in Ali Abad/Karimabad region (see Figure 9) shows a mean deviation of −0.02 m (median deviation of 0.03 m), a StdDev of 5.26 m, and an NMAD of 4.54 m. Ref. [16] reported a mean deviation of −4.1 m, a StdDev of 13.6 m, and an NMAD of 12.3 m using a DEM derived from WorldView-II and GeoEye-1 satellite images. While our results may appear better based on the numbers, it is important to note that [16] used a reference DEM of higher quality compared to the GLO-30 DEM. Therefore, such comparisons should be made cautiously, assuming, as with the PSS DEM from 2017, equivalent outcomes.

Analyzing the patch in the steep terrain north of Ali Abad/Karimabad (see Figure 10) revealed findings consistent with the 2017 dataset. However, in the southern region, no systematic intersection errors were observed. Larger intersection errors along the steep slopes of the Hunza River are likely attributable to a combination of low texture, steep incidence angles, actual terrain changes, or inaccuracies in the reference DEM, which remain uncertain.

The analysis of the Muchowar glacier patch in Figure 11 also reveals patterns similar to those observed in 2017. Two notable aspects emerge from the error map. First, the glacier appears to have undergone significant movement during the 12-day image acquisition period, explaining the pronounced intersection errors along the glacier tongue. Systematic differences are also evident in the DoD, aligning with the intersection errors observed here.

The error map also highlights systematically higher errors in the northern region. Examination of the Count map indicates a boundary between the image pairs used, suggesting that the configuration of these pairs may have been suboptimal, similar to the southern area of Ali Abad/Karimabad in 2017. Interestingly, these elevated errors are not as clearly reflected in the DoD, possibly because they remain below some kind of “Level of Detection” compared to the accuracy and GSD of the GLO-30 DEM.

The DoDs for both timestamps show mostly random errors, with larger deviations in steep shadowed areas affecting stereo reconstruction. This suggests generally accurate DEMs with some localized discrepancies due to terrain or environmental factors. A comparison in steep areas with reference data of higher accuracy than the GLO-30 DEM would be interesting for future work.

4.1.3. Surge Investigation

Figure 12 illustrates the change detection results, derived from DoD by subtracting the 2017 PSS DEM from the 2019 PSS DEM. Glacier outlines from Muchowar and Shisper glaciers, obtained from the RGI, were applied as a mask. The results closely align with the findings of [16]. The DoD and elevation profiles derived from the DEMs effectively highlight the glacier’s surge event, illustrating the significant mass transfer within the glacier area. It should be noted that performing DoD calculation with higher temporal resolution, i.e., using DEMs from more time stamps, would likely yield a more comprehensive depiction of surge dynamics, capturing seasonal variations and transient behaviors not discernible from the two temporal points utilized in this study. This further demonstrates the utility of the method in capturing dynamic changes in glacial morphology. A detailed glaciological interpretation of these results lies beyond the scope of this study. However, readers seeking a deeper understanding may refer to [37].

4.2. Bøverbrean

Bøverbrean, a small glacier situated in the southern part of Scandinavia, Norway, within the north-western area of Jotunheimen National Park’s Smørstabb Massif, is part of a cluster of small glaciers. It is a popular hiking destination for tourists and has shown significant changes in its front position, both in position and height, particularly over the last decade [34,35,41]. This glacier is continually monitored and serves as an excellent study site to assess whether Planet4Stereo can effectively detect and monitor smaller-scale changes.

4.2.1. Processing—Bøverbrean Glacier

A DEM, shown in Figure 13, was generated to represent the glacier’s condition at the end of the glaciological year in 2021. This DEM was derived from PlanetScope imagery captured between August and September of that year (see

Table 3. Settings of Bøverbrean glacier DEM processing.

	PlanetScope DEM 2021
Configuration
Number of images	10 (NIR)
Scene IDs	20210828_095603_87_2455, 20210828_100017_04_2459, 20210829_102532_1001,
	20210829_102533_1001, 20210913_100744_79_2251, 20210913_105036_37_2413,
	20210914_102653_1040, 20210914_104522_58_240f, 20210914_104524_89_240f,
	20210928_100200_0e20
Minimum overlap (%)	10
Minimum convergence angle (°)	4
Ortho workflow	yes (GLO-30 DEM)
Bundle-block adjustment
Number of triangulated points (sparse)	3386
Mean residuals of images (px)	0.6
Mean residuals of triangulated points (px)	0.6
Stereo reconstruction
Mean intersection error (m)	3.1 (max: 27.2)
DEM sampling
DEM coverage (km2)	752.2

), which exhibited minimal snow cover. The selected dataset includes ten PlanetScope scenes, illustrated in Figure 14. In particular, no images were available from an ascending orbit during the study period. It should be noted that a period of one month can already mean one meter of melting in glaciated regions, causing inconsistencies between the images, but this is far below the level of detection using PlanetScope-derived DEMs, as determined by [19], and is therefore accepted due to the low availability of usable images. The GLO-30 DEM served as approximation DEM, following a methodology consistent with the processing of the Shisper glacier.

Details of the processing parameters and statistics of the calculated DEM are provided in Table 3. It is important to note that the DEM extends beyond the Bøverbrean region itself. The numbers thus encompass the entire overlapping image area, meeting minimum overlap and convergence criteria, rather than solely focusing on the densest overlap area specific to Bøverbrean. Despite reporting a mean intersection error of 3.1 m (0.8 px), it should be noted that this average pertains to the entire processed area. Specifically around Bøverbrean (outlined in Figure 14), the mean and maximum errors are lower, at 2.8 m (0.7 px) and 10.2 m (2.6 px), respectively.

4.2.2. Evaluation

We also used demcoreg to assess the accuracy of the PlanetScope DEM by comparing it with a reference DEM in stable terrain (see Figure 15). For this purpose, we utilized a LiDAR DEM nasjonal detaljert høydemodell 2020 provided by the Norwegian authorities, acquired at the beginning of August 2020 using airborne laser scanning. This DEM has a GSD of 2 m and a vertical accuracy of 0.12 m [42]. Although coverage of the southern Smørstabb Massif is limited, the available data are sufficient to evaluate the accuracy of the PlanetScope DEM around the Bøverbrean area.

The observed errors range from ±40 m, with an NMAD of 6.47 m and a StdDev of 14.24 m. Pronounced negative deviations were noted inside the southern Smørstabb Massif, likely due to shadowed areas over steep terrain. Planetscope images, captured in the local morning, often expose strong cast shadows at such locations, as shown in Figure 15. The orthophoto mosaic reveals darkened, steep areas, both of which pose challenges for surface reconstruction due to ambiguities in image matching or steep intersection angles during triangulation.

However, it is conceivable that these offsets are also attributable to actual terrain changes as the reference dataset and PlanetScope DEM were acquired with a one-year delay. To clearly disentangle these effects, comparisons with reference data collected concurrently would be necessary.

A statistical evaluation of the DoD for the western area outside the Smørstabb Massif, near the terminus of the Boverbrean glacier (highlighted with stronger contour in Figure 15), reveals minor deviations, with mean differences of several decimeters (−0.80 m) and median differences of a few centimeters (0.04 m) compared to the reference LiDAR DEM. The area features mostly flat terrain with moderately steep sections, while the far western region shows steeper slopes and again larger deviations in the DoD.

The analysis highlights significant errors in steep shadowed areas, likely due to challenges in surface reconstruction, while flatter regions show only minor deviations. However, the potential influence of actual terrain changes remains given the one-year temporal gap between datasets, emphasizing the importance of concurrent validation data for clearer interpretation.

4.2.3. Change Detection in South Smørstabb Massif

A comparative analysis of the southern Smørstabb Massif, including Bøverbrean, was conducted to assess glacier changes by comparing a 2021 PlanetScope DEM with one acquired by the Norwegian authorities in September 2013 from aerial images using aerotriangulation. The latter DEM has a GSD of 1 m and an average vertical accuracy of 0.6 m, with maximum reported errors of 2 m in steep terrain. Although the dataset does not cover the entire extent of the area represented in the PlanetScope DEM, it nonetheless provides valuable insights into elevation changes in the southern Smørstabb Massif. The temporal gap of eight years is particularly relevant as long-term monitoring by the Norwegian Water Resources and Energy Directorate (NVE) has documented significant glacier retreat during this period [34], and thus a substantial loss of ice volume within the glacier’s delineated area is expected when comparing the DEMs.

The results of the DoD analysis are presented in Figure 16. As anticipated, the analysis reveals substantial negative elevation changes across the entire Smørstabb Massif. High positive deviations are observed in areas influenced by glacial lakes or shadowed regions as PlanetScope scenes are acquired sun-synchronously during the local morning.

Focusing on Bøverbrean, the distinct tapering shape of the glacier terminus is clearly visible and aligns well with glacier outlines from the 2000s. Although detailed information on ice thickness changes at the terminus is unavailable, the observed shifts in the glacier front position closely match the outlines documented in the RGI dataset, further supporting the validity of the detected changes. The differences in the southern part of the Smørstabb Massif show a mean and median decrease of approximately −15 m. This is plausible when considering observations reported by sources such as [34].

4.3. Comparison with Agisoft Metashape

To further evaluate the precision of DEMs processed using Planet4Stereo, we compared it with Agisoft Metashape (version 2.0.2), a leading Structure-from-Motion tool supporting RPC processing since version 1.5.0. Using Metashape’s standard pipeline, we processed the 2019 Shisper glacier dataset with default settings, performing image alignment at “high” accuracy and generating the 3D terrain model from depth maps. The final model was exported in ellipsoidal coordinates (WGS84, EPSG:4326). This study also references [16], where Metashape was employed for DEM reconstruction with pre-optimized RPCs. We focus on comparing our pipeline, which incorporates RPC optimization via ASP, to Metashape as a closed solution. Figure 17 illustrates a DoD calculation against the GLO-30 DEM, following methods outlined in Section 4.1.

The qualitative analysis reveals that the DEM generated using Metashape appears more complete, likely due to fewer restrictions in image pair selection and potentially superior matching algorithms. Additionally, the Metashape DEM exhibits finer detail despite comparable resolution, possibly attributed to differences in the parameters, such as window size, for dense image matching in DEM reconstruction.

While the DoD comparison shows similar results between the Metashape and Planet4Stereo DEMs, notable discrepancies arise in specific regions. For instance, in steep terrain near the terminus of the Muchowar glacier, the Metashape DEM performs slightly worse. In the stable area of Ali Abad, Metashape reconstructs a region in the southwest exhibiting negative deviations. This anomaly could stem from an image pair with poor stereo capability, which was pre-filtered in Planet4Stereo’s pipeline but included by Metashape due to its lack of preselection.

In conclusion, while the Metashape DEM qualitatively appears more complete, its accuracy is not necessarily superior, with completeness influenced by differences in image pair filtering.

5. Conclusions

In this article, we presented an approach for efficiently generating DEMs of landscapes using an easy-to-use open-source pipeline called Planet4Stereo. This method enables the processing of multi-temporal, inherently non-stereo, and cost-effective PlanetScope imagery. Our experiments demonstrate that the results produced by the pipeline are comparable in quality to those described in the literature, which often rely on proprietary or less-accessible tools. The pipeline empowers researchers in fields such as geomorphometry and glaciology to derive valuable insights into landscape deformation with high spatiotemporal resolution without requiring extensive expertise in photogrammetry. The core of the pipeline, ASP, is continuously evolving and integrating new methods that could further enhance stereo matching capabilities. In Planet4Stereo, only minor adjustments to the script (e.g., changing the matching method flag) are necessary in this case. At the same time, Planet Labs is developing new generations of satellites. Future work could focus on both the technical advancement of Planet4Stereo, incorporating updates from ASP, and evaluating its suitability for data from upcoming satellite generations.

Furthermore, key components of the pipeline—particularly the stereo matching and triangulation stages—could be substituted with corresponding modules from other open-source tools, such as MicMac [43], to assess potential improvements in disparity estimation accuracy and surface reconstruction efficiency.

Several limitations warrant discussion. The current approach relies on approximation data for co-registering the pairwise-generated PSS DEMs, which may introduce biases in areas where the approximation DEM is less accurate or outdated. This issue is particularly pronounced in dynamic environments, such as fast-moving glaciers, where rapid surface changes challenge co-registration accuracy. Integrating multi-source datasets, such as Sentinel-1 synthetic-aperture radar (SAR) data (GSD of 20 m, six-day revisit), may enhance robustness by providing approximation data with minimal temporal offset to the used PlanetScope images. Additionally, SAR data could help to fill in gaps in PlanetScope-derived DEMs, such as in shadowed or cloud-covered areas, although resolution differences necessitate extrapolation. While SAR-derived DEMs have lower spatial resolution, their resilience in adverse conditions and ability to capture dynamic changes could improve co-registration and overall DEM quality. However, effective interferometric SAR (inSAR) processing requires a comprehensive understanding of, e.g., baseline geometry and phase decorrelation [44].