Century-Scale Earth Observation: Systematic Review of Georeferencing Methods for Historical Aerial and Satellite Imagery

Highlights

What are the main findings?

• Despite advances in algorithmic processing, georeferencing historical images remains bottlenecked by a heavy reliance on manual and partially automated workflows due to inherently incomplete metadata and uncertain acquisition geometries.
• Ground control points persistently dictate the geometric constraints of historical imagery, exposing a critical void in fully automated methodologies that operate independently of these points and can adapt to complex landscape changes.
• Accuracy assessment practices are fundamentally inconsistent across the literature, as most studies report control-point residuals rather than independent checkpoint validation, and reporting conventions range from quantitative RMSE to qualitative visual inspection, critically undermining cross-study comparability and the cumulative scientific value of georeferenced historical archives.

What are the implications of the main findings?

• A New Methodological Paradigm: Georeferencing historical data can no longer be treated as a simplified secondary stage of modern photogrammetry. These findings imply that it should be recognized as a distinct methodological domain requiring workflows designed specifically for structural uncertainty, sensor heterogeneity, and severe temporal landscape transformations.
• A Shift in Validation and Reporting: The current inconsistency in accuracy assessment is a massive liability. The implication here is that the field must pivot toward evaluations that account for uncertainty and validation driven by specific applications. We need mandatory transparent workflow reporting that relies on independent checkpoints rather than just residual values from control points, ensuring datasets are fit for their intended scientific purpose.
• As the first PRISMA-compliant synthesis of historical georeferencing practices, this review establishes an empirical baseline that can directly inform the development of standardized validation protocols, guide the integration of emerging deep learning and vision foundation model approaches, and prioritize capacity building in geographically underrepresented regions where pre-satellite environmental baselines are most urgently needed.

Abstract

Historical remote sensing imagery, including archival aerial photographs and declassified satellite imagery, has been increasingly used to extend earth observation records into periods not covered by modern satellite missions. However, the broader application of these data remains constrained by georeferencing challenges related to incomplete metadata, uncertain acquisition geometry, and heterogeneous image characteristics. This systematic review examines georeferencing practices for historical remote sensing imagery. Out of the 2547 studies identified in the literature, 205 peer-reviewed journal articles were deemed eligible for analysis. This systematic review provides the first comprehensive, PRISMA-compliant synthesis of georeferencing practices for historical remote sensing imagery, analyzing 205 peer-reviewed studies to establish methodological patterns and identify critical gaps. The review considers imagery types, spatial and temporal distributions of case studies, georeferencing workflows, geometric constraints, and accuracy reporting practices. The results indicate a strong reliance on ground control points and a clear preference for manual or semi-automatic georeferencing approaches, while fully automatic methods remain rare. Although the use of historical imagery has increased over time, its potential has not been fully exploited due to persistent georeferencing difficulties, and study areas are often spatially limited or selectively processed to achieve acceptable accuracy. Nevertheless, properly georeferenced historical imagery has been widely applied to land-cover analysis, geomorphology, cryosphere research, hazard assessment, and archeology by extending observation records into earlier decades.

1. Introduction

Over the past several decades, earth observation research has relied primarily on satellite-based remote sensing datasets to characterize land surface conditions and their changes at regional to global scales [1,2]. In practice, the satellite records utilized for quantitative and comparative environmental studies are generally limited to the post-1980s era, following the advent of consistently processed and radiometrically reliable Landsat imagery [3,4,5]. Consequently, land surface conditions from earlier periods remain underrepresented in standard earth observation analyses [6,7]. This limitation makes it difficult to evaluate contemporary environmental change against earlier baseline conditions or to place current observations within a longer historical context [7,8]. Extending spatial observations into earlier decades is therefore critical for several Earth observation applications, including the establishment of pre-satellite environmental baselines, long-term land-use and land-cover change analysis, and the monitoring of gradual environmental processes such as glacier retreat, wetland loss, and coastal erosion, which often unfold over timescales exceeding the available satellite record. To bridge this observational gap, historical remote sensing imagery, including early satellite photographs and archival aerial images, provides essential spatial information required to extend the observation record into the pre-satellite era [9,10].

Historical remote sensing imagery was acquired using a variety of platforms and sensors prior to the widespread adoption of modern digital satellite systems, resulting in notable differences in image characteristics and acquisition geometry [11]. These historical datasets can be broadly categorized into archival aerial photographs, early civilian satellite imagery, and declassified reconnaissance satellite photographs, reflecting successive stages in the development of remote sensing technology [12,13]. Archival aerial photographs often extend back to the early twentieth century and provide high spatial detail suitable for fine-scale landscape analyses [14,15]. Early civilian satellite missions expanded observational capabilities beyond the spatial limitations of aerial surveys, while declassified reconnaissance imagery further enabled near-global coverage at high spatial resolutions [16,17,18]. Collectively, these diverse data sources extend spatial observation coverage into periods otherwise inaccessible via modern satellite missions [13,15,19,20]. Despite their immense potential for long-term monitoring, these historical datasets present significant technical challenges [21,22]. In particular, inherent differences in sensor characteristics and acquisition conditions introduce substantial geometric inconsistencies that complicate their seamless integration with contemporary geospatial datasets [23,24].

During the early stages of remote sensing development, imaging technologies, sensor calibration, and geometric modeling had not yet fully matured, and standardized acquisition protocols were often lacking [25,26]. As a consequence, many historical datasets were collected without comprehensive or consistently documented metadata, including detailed sensor specifications, platform orientation, acquisition geometry, and accurate ground reference information [27,28]. These limitations are further compounded by the use of film-based sensors, non-nadir viewing configurations, and heterogeneous acquisition conditions across platforms and time periods, which introduce complex and spatially variable geometric distortions [29,30,31]. In contemporary earth observation workflows, these factors introduce significant uncertainty in the spatial positioning and geometric consistency of features derived from historical imagery [32]. In modern photogrammetric and remote sensing workflows, comparable geometric uncertainties are typically resolved through rigorous reconstruction methods such as aerial triangulation, bundle adjustment, and increasingly through direct georeferencing systems integrating GPS–IMU or LiDAR technologies [33,34,35]. These approaches rely on well-documented sensor models, calibration parameters, and accurate positioning information [36,37]. However, for many digitized historical photographs, camera orientation parameters and other acquisition information are no longer available [38,39]. In such cases, georeferencing cannot rely on a complete photogrammetric sensor model and instead must be performed by estimating a geometric transformation between image coordinates and ground reference data [40,41]. From the perspective of image registration theory, this process can be viewed as aligning images acquired at different times or by different sensors through the estimation of a spatial transformation between corresponding features [42,43]. Although the underlying geometric principles remain consistent with established photogrammetric practices, the absence of complete sensor metadata and the heterogeneous characteristics of historical imagery often make the georeferencing workflow more complex in practice. Consequently, the direct application of modern photogrammetric reconstruction frameworks to historical datasets is often infeasible. This limitation complicates the integration of historical imagery with contemporary geospatial datasets [44,45]. Under these conditions, the effective use of historical imagery depends critically on establishing reliable spatial reference and geometric alignment, particularly through retrospective georeferencing strategies tailored to incomplete archival data [20,46].

In historical remote sensing applications, georeferencing assigns archival imagery an explicit spatial reference and enables its integration into spatially explicit Earth observation analyses. It is typically performed as the initial step in the processing workflow, particularly when complete sensor models, calibration reports, or direct positioning information are unavailable [47,48]. Through this process, relationships between image coordinates and real-world geographic locations are established, allowing geometric distortions to be corrected and historical images to be positioned within a common spatial reference framework [49,50]. This alignment enables observations acquired at different times, from different platforms, and with different sensors to be compared and combined with modern geospatial datasets [51,52]. It is critical for long-term Earth observation applications, as georeferencing uncertainty can bias land-use and land-cover change analysis, environmental monitoring, and the reconstruction of pre-satellite baselines. Unlike comprehensive photogrammetric reconstruction, which relies on fully parameterized sensor models and orientation solutions, georeferencing in historical remote sensing studies generally involves the post hoc alignment of archival imagery to contemporary coordinate reference systems using ground control points, feature-based matching, or transformation models [8,53]. In practice, georeferencing can be particularly challenging for early aerial photographs and analogue imagery that have been digitized from archival film [22,54,55]. In such cases, metadata may be incomplete, unavailable, or poorly preserved, and sensor geometry may be poorly documented or unknown [56,57]. Additional geometric distortions may also arise during film digitization or scanning processes, particularly when non-photogrammetric scanning methods are used [40,58,59]. Consequently, georeferencing workflows often differ substantially depending on image type and data availability, including differences in assumptions, reference data sources, and levels of user intervention [48]. Such variations in spatial accuracy can propagate into downstream analyses, potentially biasing change detection, area estimates, and long-term environmental reconstructions [60]. The choice and rigor of georeferencing methodology is therefore not merely a preprocessing consideration but a determinant of scientific validity in long-term Earth observation studies. Despite the increasing use of historical imagery, georeferencing practices remain highly heterogeneous across case studies and imagery types [8]. Most studies are application-driven and context-specific, with limited systematic comparison of workflow configurations, geometric constraints, automation levels, and accuracy reporting approaches [4,61,62]. As a consequence, methodological patterns and commonly reported limitations are often discussed only within individual case studies [53,56,57]. Systematic synthesis of georeferencing practices across different imagery types and application contexts remains limited.

In this study, we present a comprehensive review of georeferencing approaches for historical remote sensing imagery, focusing on how spatial challenges specific to historical data have been addressed in existing research. The review considers a wide range of historical imagery types and acquisition contexts, highlighting how differences in image characteristics and data availability influence georeferencing strategies. Existing approaches are organized according to their methodological foundations, with emphasis on distinctions in assumptions, reference data usage, automation levels, and achievable geometric accuracy. Common sources of uncertainty and methodological limitations are examined to clarify how these factors affect the reliability of subsequent earth observation analyses. This review aims to provide a structured synthesis of existing georeferencing practices for historical remote sensing imagery and to inform methodological choices for robust spatial integration of archival imagery, and ultimately to support more reliable reconstruction of long-term environmental change, pre-satellite baseline establishment, and multi-temporal land surface analysis.

2. Methods

This systematic literature review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [63]. The review protocol was not registered. This review was conducted following a structured systematic review framework [64,65]. A scientific literature database was queried, and explicit eligibility criteria were established prior to study selection. Identified records were subsequently screened and assessed against these criteria, after which the included studies were subjected to a structured analytical procedure. The PRISMA 2020 [63] framework was used to guide transparent reporting of the review process, ensuring that study identification, selection, and synthesis decisions were documented in a clear and reproducible manner. An overview of the review workflow and study selection process is provided in Figure 1.

2.1. Search Strategy and Data Sources

Our study used the Web of Science Core Collection as the sole bibliographic database. This database was selected because it provides strong coverage of peer-reviewed journals in remote sensing, photogrammetry, geosciences, and related interdisciplinary fields, while also offering standardized metadata and Application Programming Interface (API)-based access that support transparent and reproducible literature retrieval. The literature search was conducted programmatically through the Web of Science API within the R environment to ensure a systematic and reproducible retrieval process. The search strategy was designed to identify studies addressing historical remote sensing imagery in conjunction with georeferencing and geometric correction procedures. Article titles, abstracts, and author keywords were queried using a structured search composed of two thematic components. The first component targeted historical, aerial, and declassified imagery, including terms related to historical imagery, historical aerial photographs, declassified imagery, and major declassified satellite missions (CORONA, ARGON, LANYARD, HEXAGON, and GAMBIT). The second component focused on georeferencing and orthorectification procedures and included terms related to georeferencing, orthorectification, orthophotos, and orthoimages. These two components were combined using the logical operator AND to ensure that retrieved records addressed both historical imagery and georeferencing-related aspects. Wildcard symbols (*) and quotation marks were applied where appropriate to capture variant word forms and reduce ambiguous matches identified during preliminary searches. The search was limited to peer-reviewed journal articles indexed in the Web of Science, with no lower limit on publication year and an upper limit of 2025. The complete Web of Science search query and automated retrieval procedure are provided in Supplementary File S1, which contains the exact query syntax used for literature retrieval through the Web of Science API. Duplicate record identification and removal were performed using a dedicated R script provided in Supplementary File S2, ensuring transparency and reproducibility in data preparation. Although this strategy supported retrieval consistency and reproducibility, it may have excluded relevant studies not indexed in the Web of Science Core Collection.

2.2. Study Screening and Eligibility

After retrieval, all records were screened in a stepwise manner to identify studies relevant to the scope of this review. First, duplicate records were identified and removed using an R-based deduplication procedure. Second, titles and abstracts were screened against predefined eligibility criteria. At this stage, records were excluded if they did not involve historical remote sensing imagery, did not address georeferencing, orthorectification, or related geometric correction procedures, or were clearly outside the methodological scope of this review. The full texts of the remaining studies were then assessed for eligibility. Full-text screening focused on whether the study explicitly used historical aerial imagery or declassified satellite imagery and whether the georeferencing or orthorectification procedure was described in sufficient methodological detail to support extraction and coding. Studies were excluded at this stage if they relied only on already georeferenced image products, lacked direct methodological discussion of georeferencing procedures, or did not provide sufficient information to support structured coding. The final included studies therefore comprised publications that explicitly addressed the methodological practice of georeferencing historical remote sensing imagery. Screening was conducted by one reviewer using predefined eligibility criteria. To improve consistency, the eligibility criteria were formalized and pilot-tested on a subset of records prior to full screening, which helped clarify the application of inclusion and exclusion rules before the main screening procedure began. Uncertain cases were re-examined through repeated full-text checking, and decisions on borderline studies were made in consultation with co-authors before final inclusion decisions were reached.

2.3. Data Extraction and Coding Framework

We systematically coded each included study using a structured extraction protocol designed to capture the main contextual and methodological characteristics of the reviewed studies. The coding framework was organized into several analytical domains, including study context, georeferencing workflow characteristics, geometric constraint structures, and broader research applications and temporal patterns represented in the literature. The study context domain recorded the types of historical imagery used, the geographic distribution of case studies, and the temporal coverage of the datasets. The georeferencing workflow domain documented the geometric transformation models employed and the reported level of automation in the georeferencing procedures. Automation levels were coded as manual, semi-automatic, or automatic according to the extent to which ground control point identification and transformation estimation were performed by human operators or supported by algorithmic procedures. The geometric constraint domain captured the constraints used to estimate geometric transformations, including ground control points (GCPs), digital elevation models (DEMs), and sensor metadata. The application and temporal domain recorded the research application areas of historical imagery and the publication year of each study to examine temporal patterns in the literature.

Data extraction and coding were conducted by one reviewer using a structured extraction table and a predefined coding framework. To improve consistency, the coding framework and variable definitions were established prior to full coding. Ambiguous coding cases were revisited through repeated full-text checking and discussed with co-authors before final coding decisions were made. Variables were recorded only when explicitly reported or directly identifiable in the source studies. When information was missing, ambiguous, or insufficiently specified, conservative coding rules were applied and such information was recorded as absent rather than inferred. Formal risk-of-bias assessment tools were not applied because this review did not aim to estimate effect sizes or compare quantitative outcomes, but rather to synthesize methodological practices in historical image georeferencing. Given the substantial heterogeneity in imagery types, reference data availability, validation strategies, and reporting conventions, standardized bias scoring was not considered appropriate for this review. Instead, potential sources of bias and uncertainty were addressed through conservative coding and structured comparative synthesis. All coded data were managed in structured databases and analyzed in R (version 4.5.1). Visualizations were produced using ggplot2, UpSetR for constraint pattern analysis, and networkD3 for Sankey diagrams [66,67,68]. ggplot2 was used to generate standard statistical graphics, including bar charts and temporal distribution plots. UpSet diagrams were used to represent intersection patterns among multiple geometric constraint types, allowing complex combinations of constraints reported in the literature to be displayed clearly. Sankey diagrams were used to visualize the distribution and linkage relationships between imagery sources and research application domains. In this context, the Sankey diagram represents categorical linkages rather than temporal flows, with the width of each connection indicating the number of studies associated with each imagery–application pairing.

2.4. Study Selection Results

The literature search conducted in the Web of Science database using the predefined queries returned a total of 2547 peer-reviewed journal articles. Following the initial retrieval, duplicate records were identified and removed, resulting in 2546 unique articles for subsequent screening. These records were then screened based on their titles, abstracts, and author keywords to assess their relevance to historical remote sensing imagery and georeferencing-related methodologies. After the initial screening, 352 articles were retained for full-text assessment. A detailed examination of the full texts led to the exclusion of studies that did not involve historical imagery, lacked explicit descriptions of georeferencing or orthorectification procedures, or provided insufficient methodological detail. Ultimately, 205 articles met all inclusion criteria and were retained for analysis in this review. The final corpus consists exclusively of peer-reviewed journal articles that directly address the georeferencing of historical remote sensing imagery.

3. Results

3.1. Empirical Context of Historical Imagery Georeferencing

Across the reviewed corpus, historical imagery used for georeferencing falls primarily into two categories: historical aerial imagery and declassified satellite imagery such as CORONA, ARGON, and HEXAGON [69,70,71]. Figure 2 summarizes the usage shares of these imagery types across the reviewed studies. Historical aerial imagery constitutes the majority with 174 studies, accounting for 83.9% of the reviewed cases, whereas declassified satellite imagery contributes 33 studies, representing 16.1%. Within the declassified subset, CORONA dominates with 22 studies (66.7%), followed by ARGON with 6 studies at 18.2% and HEXAGON with 5 studies (15.1%). Overall, the results show that most reported georeferencing practices rely on aerial photograph archives, which constitute the dominant imagery source across the reviewed studies [22,72,73,74]. By contrast, declassified missions represent a smaller portion of the reviewed corpus. Nevertheless, their importance extends beyond their numerical share because they provide systematic, national-scale historical observations, broad spatial coverage, and access to critical pre-Landsat imagery. These features make declassified satellite imagery particularly valuable for extending Earth observation records and supporting long-term historical landscape reconstruction.

The reviewed studies exhibit broad geographic coverage, with study areas spanning multiple continents (Figure 3). The United States constitutes the largest national share with 38 studies [50,75], followed by Italy with 24 [76,77] and Spain with 14 [78,79]. Beyond these leading contributors, the spatial distribution reveals a pronounced concentration across Europe [80,81,82] and North America [83], with sparser representation across Asia [84,85], Oceania [86,87], and parts of Africa [88,89]. Coverage extends into high-latitude environments [14,90], confirming that historical imagery georeferencing has been applied across a wide range of geographic and climatic settings beyond the mid-latitudes. Nevertheless, many countries contribute only one or two studies, and tropical regions and much of the Southern Hemisphere remain substantially underrepresented. The evidence base is therefore global in scope but markedly uneven in geographic density. This imbalance likely reflects, at least in part, differences in data availability, archival access, and the extent of digitization efforts across regions. Europe and North America generally have longer-established aerial survey programs, more accessible archival infrastructures, and broader availability of supporting reference data, all of which lower the practical barriers to historical image georeferencing. By contrast, the limited representation of Asia, Africa, and tropical regions suggests that current methodological patterns may be disproportionately shaped by evidence from data-rich, temperate settings. This uneven representation has important implications for interpretation, because the workflow choices, constraints, and reported performance documented in the literature may not fully capture the challenges of georeferencing historical imagery in underrepresented regions with different environmental conditions, archival histories, and reference data limitations.

The temporal footprint of imagery used in the reviewed studies spans more than a century, extending from <1900s to the 2020s (Figure 4). Figure 4 presents decade-level coverage at the study scale and shows which decades of imagery each paper incorporated. The distribution is anchored in pre-1980 decades, with the 1930s–1960s appearing most frequently across studies [91,92]. Earlier decades are comparatively rare [93,94,95]. The <1900s category occurs only sporadically, and the earliest reported imagery dates back to the 1850s [96]. Later decades after the 1980s appear less consistently in the matrix and mainly extend the record forward into the 2000s–2020s [97,98]. At the level of study design, multi-decade configurations are common, and the decade blocks are often separated by gaps, indicating that many studies assemble selective temporal snapshots rather than a continuous decade-by-decade record [99,100,101]. Beyond the decade aggregation shown here, many papers report imagery dates at finer granularity, such as specific acquisition years or multiple image dates within the same decade, which underlies the decade-level synthesis used in this review [102].

3.2. Geometric Transformation Models and Geometric Constraints

Across the reviewed literature, three types of geometric transformation models are reported in historical image georeferencing (Figure 5): physical sensor models, empirical transformation models, and hybrid implementations that combine elements of both approaches. These models define the mathematical relationship used to transform image coordinates into ground-based coordinate systems [103,104]. Physical sensor models represent the most frequently reported category, appearing in 102 studies [105,106]. These models describe image geometry through physically based sensor parameters and photogrammetric formulations. Empirical transformation models constitute the second most common category and are documented in 85 studies [107,108]. In these approaches, the transformation between image and ground coordinates is estimated directly from ground control points without explicitly reconstructing the original sensor geometry. Hybrid implementations that combine physical sensor modeling with empirical transformations are reported in only 18 studies [56,109].

In addition to transformation models, the reviewed studies report several types of geometric constraints used to estimate these transformations (Figure 6). The following analysis summarizes the constraint information explicitly reported in the reviewed studies and extracted through the coding procedure. Three main types of geometric constraints were identified in the reported literature: ground control points (GCPs), digital elevation models (DEM), and sensor metadata. Among the reported cases, GCPs appear in 169 studies, DEM in 107 studies, and sensor metadata in 93 studies [110,111,112]. In addition, 19 studies did not report sufficient information to classify the constraint structure [113,114]. The reported constraints occur in several combinations across individual georeferencing workflows (Figure 6). The most frequently reported constraint combination integrates GCPs, DEM, and sensor metadata and is documented in 71 studies [115,116,117]. A second commonly reported configuration relies on GCPs only and occurs in 62 studies [118,119]. Other reported combinations include GCPs with DEM (20 studies) and GCPs with sensor metadata (16 studies) [120,121,122]. Configurations reported without GCPs are comparatively uncommon, including DEM with sensor metadata (5 studies), DEM-only configurations (11 studies), and sensor-metadata-only implementations (1 study). These counts summarize the constraint structures reported in the reviewed papers rather than complete operational workflows, because some studies did not fully specify all constraint sources used during georeferencing [123,124]. In several studies classified as DEM-only, DEMs were explicitly reported as terrain reference data, whereas other potential constraints were not described in sufficient detail to be coded [125,126,127].

3.3. Utilization of Georeferencing Outcomes Across Research Themes

Accuracy validation practices vary substantially across the reviewed studies. Positional error metrics derived from control information are frequently reported, with root mean square error and related measures appearing most often [128,129,130,131]. However, the basis of these metrics differs across studies. In many cases, reported errors correspond to residuals at ground control points used during georeferencing, whereas fewer studies report accuracy estimates derived from independent check points [54,132]. Because residuals at ground control points are calculated using the same points that constrain the georeferencing solution, they typically represent internal fitting errors rather than independent validation of positional accuracy [133,134,135]. Consequently, residual-based metrics may underestimate the true spatial error compared with validation based on independent check points [136,137,138]. This difference in validation strategy limits the direct comparability of reported positional accuracy values across studies [139,140]. Beyond differences in validation strategy, reporting practices also vary widely across the literature. Alongside numerical metrics, qualitative reporting is also common, including visual inspection of spatial alignment with reference datasets or brief descriptive statements confirming acceptable correspondence [52,141,142]. Differences are also evident in the level of detail with which accuracy information is reported. Some studies report numerical error values together with descriptions of the geometric information used during georeferencing, whereas others provide limited numerical detail or rely primarily on qualitative confirmation of spatial alignment [143,144]. Reporting detail also appears to be associated with workflow complexity. Studies employing multiple sources of geometric information, such as combinations of ground control points, digital elevation models, and camera models, often report accuracy information alongside these inputs [53,57,145]. In contrast, studies based on simpler configurations more frequently omit explicit numerical metrics or restrict reporting to brief descriptive statements [61,146].

Figure 7 presents a Sankey diagram showing the categorical mapping of reviewed studies between data sources and scientific applications. It provides an overview of how different imagery types are distributed across research themes. Methodological studies constitute the largest category, accounting for 37 studies (18% of the reviewed literature). These studies primarily focus on the technical processing of historical imagery itself, particularly the development and evaluation of georeferencing approaches and related preprocessing workflows. Among the application-oriented studies, coastal research represents the largest thematic group, accounting for over 15% of the total (n = 34). In contrast, geomorphological and fluvial studies occur less frequently, with 18 and 15 studies, respectively, each accounting for less than 10% of the reviewed literature [55,147,148]. Several other research domains show broadly comparable study volumes. Ecosystem-related and cryosphere studies each account for nearly 14% of the reviewed literature, with 29 and 28 studies, respectively [149,150,151]. Land-use and land-cover studies account for 11.2% of the reviewed literature, while archeological applications represent 10.2% [152,153,154]. Imagery types also differ in their distribution across research domains. Historical aerial imagery appears across a wide range of research themes, including coastal, ecosystem, land-use, and geomorphological studies [155,156,157]. Declassified satellite imagery is represented across multiple research domains, but most studies are concentrated in methodological research [158,159,160]. Archeological and cryosphere-related studies also account for a notable share, whereas applications in other domains remain limited [161,162]. No studies based on declassified satellite imagery were identified in fluvial research, and only single-digit studies were identified in coastal, land-use, and ecosystem-related applications.

3.4. Temporal Trends in Georeferencing Practice

The reviewed studies span the period from 1999 to 2025 (Figure 8), with a total of 205 studies included. Overall, the number of publications exhibits a clear increasing trend over time. Prior to 2010, annual publication counts remained low, with most years reporting fewer than five studies and isolated peaks in approximately 2002, 2006, and 2009. From 2010 onward, annual publication counts increased steadily, with a more pronounced rise after 2015, when annual counts consistently exceeded ten studies. A substantial share of the reviewed literature was published between 2018 and 2023. This shift from isolated peaks to sustained publication output after 2015 reflects a consolidation of research activity in historical image georeferencing.

The temporal distribution of georeferencing workflow automation levels is summarized across five-year publication intervals in Figure 9. In the earliest interval (1995–2000, n = 3), all studies used manual workflows. Semi-automatic workflows first appeared in the 2001–2005 interval (n = 16), representing approximately 25% of studies in that period, while manual workflows remained dominant. In the 2006–2010 interval (n = 31), the share of semi-automatic workflows increased to approximately 40%, approaching parity with manual workflows. Fully automatic workflows were not reported. From 2011 to 2015 (n = 31), semi-automatic workflows became the most frequently reported approach, exceeding manual workflows. Fully automatic workflows appeared for the first time, accounting for fewer than 5% of studies. In the 2016–2020 interval (n = 62), semi-automatic workflows dominated, accounting for more than 60% of studies. Manual workflows declined to approximately 35%, while automatic workflows remained below 5%. In the most recent period (2021–2025, n = 62), semi-automatic workflows remained the most common (approximately 50%). However, manual workflows regained a substantial share (approximately 45%), resulting in a more balanced distribution. Fully automatic workflows, first reported from 2015 onward, consistently accounted for a small minority across all subsequent periods [22,53,56,57].

Across the full reviewed corpus, semi-automatic workflows were the most frequently reported configuration (n = 112), followed by manual workflows (n = 89) (Figure 10). Fully automatic georeferencing pipelines were reported in only four studies (Figure 10), all based on declassified satellite imagery [163,164]. This distribution indicates that, over the period 1999–2025, human involvement remained integral to most georeferencing workflows. Most workflows retained manual steps in GCP collection, correspondence verification, mismatch removal, and final transformation implementation [165,166,167]. The prevalence of GCP-based constraint structures, often combined with DEM integration and sensor metadata, is consistent with this pattern of partial rather than end-to-end automation [53,56]. The increased share of manual workflows in the 2021–2025 period suggests that operator verification remains important when processing imagery with variable radiometric quality or incomplete metadata [22].

4. Discussion

4.1. Structural Characteristics of the Historical Image Georeferencing Problem

The literature consistently indicates that georeferencing historical remote sensing imagery constitutes a different problem from that encountered in modern remote sensing workflows [168,169,170]. This difference arises primarily from the inherent incompleteness of historical imagery at the data level [171,172]. Contemporary satellite systems are designed as standardized measurement platforms, in which sensor geometry, platform position, acquisition time, and calibration parameters are explicitly recorded and distributed as integral components of the data product [173,174]. These characteristics provide a well-defined geometric basis for systematic spatial correction and integration. Historical imagery, in contrast, whether derived from early aerial photography or declassified reconnaissance satellites, was typically acquired to meet short-term reconnaissance or mapping objectives rather than long-term geospatial interoperability [175,176].

A defining feature of historical image georeferencing is the presence of multiple, interrelated data constraints that shape the range of feasible solutions. Historical images are frequently associated with incomplete or inconsistent metadata, limiting the reliable identification of acquisition parameters such as platform trajectory, camera configuration, image orientation, and timing [177,178]. In addition, many historical datasets originate from non-metric cameras or film-based imaging systems whose internal geometry was neither stable nor systematically documented, rendering assumptions commonly used in modern sensor models inapplicable [179,180]. Temporal separation further exacerbates these constraints, as landscapes recorded decades earlier often exhibit substantial natural or anthropogenic transformation, reducing the availability of persistent features that can support unambiguous spatial correspondence [181,182,183]. Historical imagery also exhibits pronounced platform heterogeneity, reflecting technological and institutional differences across acquisition periods [169,171]. These constraints should therefore be understood not merely as sources of error, but as conditions that fundamentally shape how historical imagery can be georeferenced.

Within this framework, the absence of a single standardized georeferencing solution should be interpreted as an intrinsic characteristic of the problem rather than as a limitation of existing methodologies [173,184]. The form of georeferencing that can be applied is constrained by the type and completeness of the available information. When acquisition geometry cannot be reconstructed explicitly, spatial alignment must instead be inferred from external correspondences between image space and ground coordinates. Because the informational basis of historical imagery varies substantially across sensors, time periods, and regions, no universally prescriptive georeferencing formulation can be assumed [185,186].

These structural constraints can be interpreted within the broader theoretical framework of image registration, particularly the trade-offs that arise under conditions of incomplete or uncertain information [130,131]. The tension between feature-based strategies (e.g., ground control point–based correspondence) and area-based or intensity-based matching reflects the classical correspondence problem, in which spatial alignment must be inferred from ambiguous or weakly constrained evidence [148]. Temporal radiometric instability, pronounced platform and sensor heterogeneity, and uncertain acquisition geometry violate key assumptions underlying many contemporary registration frameworks developed for modern multi-view or multi-temporal sensor data [84,108]. As a result, methods designed for well-characterized acquisition geometries and radiometrically stable observations cannot be directly transferred to historical imagery without substantial methodological adaptation [55,130]. Historical image georeferencing should therefore be understood as a distinct registration problem domain, rather than as a simplified or incomplete variant of modern georeferencing practice [172,177,187].

4.2. Workflow Practices in Historical Image Georeferencing

The predominance of manual and semi-automatic workflows in historical image georeferencing reflects a pragmatic accommodation to the conditions under which historical imagery is used, rather than a lag in methodological development [188,189]. Unlike contemporary remote sensing data, historical imagery rarely provides the complete geometric and metadata information required to support fully automated georeferencing [176,190]. In practice, incomplete and uncertain control information necessitates workflows that remain flexible and iterative, allowing key decisions to be revised as constraints are progressively evaluated [62,191]. Under these conditions, manual and semi-automatic procedures have emerged not as provisional solutions, but as stable forms of practice capable of operating reliably in heterogeneous and weakly constrained settings [185,192].

Within such workflows, human intervention functions primarily as a mechanism of validation and quality control, enabling continuous assessment of spatial plausibility, internal consistency, and contextual coherence [170,177,193]. Images of the same location acquired at different times may differ substantially in landscape structure, land use, and radiometric appearance. These differences complicate the identification of corresponding features and can cause automated matching techniques to produce incorrect correspondences [8,53]. Historical image georeferencing therefore cannot be treated as a simple GIS operation in which several control points are selected and a transformation model is accepted because the software produces an aligned image [56,57]. Instead, the georeferencing process requires explicit consideration of how the selected geometric transformation represents the relation between image coordinates and ground coordinates. When the transformation model is chosen without sufficient understanding of the imaging geometry, the resulting fit may satisfy the selected control points while still introducing spatial distortion elsewhere in the image. This risk is particularly evident when higher-order polynomial transformations are used in GIS-based georeferencing workflows. Such transformations, often described as “rubbersheeting”, can force local agreement at control points even when the underlying geometric relationship between image and ground space is poorly constrained. For this reason, reliable georeferencing requires critical evaluation of the transformation model, the reliability of control points, and the geometric consistency of the resulting spatial alignment.

The resulting diversity of georeferencing workflows should be understood as an adaptive response to structural variability in historical image sources rather than as evidence of methodological fragmentation [181,182]. Differences in image provenance, acquisition conditions, landscape transformation, and available reference information preclude the universal application of a single standardized workflow [186,194]. As a result, georeferencing practices are shaped by case-specific trade-offs among data constraints, analytical objectives, and acceptable levels of human intervention [195,196,197]. This situational dependence limits strict replication across studies but enhances robustness across diverse historical contexts. From this perspective, workflow heterogeneity represents a pragmatic form of consistency, in which practices converge not on uniform procedures, but on shared principles of adaptability, iterative validation, and interpretive control.

4.3. Implications for Analytical Reliability and Reuse

The variability in georeferencing procedures observed across studies have direct implications for the analytical usability of resulting spatial products [198,199,200]. Differences in reference data selection, transformation assumptions, and validation strategies produce outcomes whose reliability is strongly context dependent [201,202,203]. Consequently, datasets reporting comparable positional accuracies may differ substantially in their suitability for spatial analysis, reducing cross-study comparability and limiting reuse beyond the original application context [204,205,206].

These limitations are further reinforced by prevailing reporting conventions. Georeferencing accuracy is commonly summarized using a small number of global metrics, with limited attention to spatial error variability or uncertainty structure [207,208]. Such simplified reporting obscures the conditional nature of accuracy and can convey an inflated sense of precision, making it difficult for subsequent users to assess fitness for purpose or to anticipate limitations in downstream analyses [201,209].

At the same time, a persistent disconnect can be observed between reported georeferencing accuracy and analytical suitability [210,211]. Many typical applications of historical imagery are more sensitive to local distortions and relative spatial consistency than to average positional error alone [212,213,214]. When evaluation metrics are not aligned with these analytical sensitivities, reported accuracy provides limited guidance on whether a georeferenced product can reliably support subsequent spatial analyses.

From this perspective, improving the reliability of georeferenced historical imagery does not primarily depend on increasing methodological complexity, but on a shift in evaluation and reporting emphasis [215,216,217]. Greater attention to uncertainty-aware reporting, application-oriented accuracy interpretation, and robustness with respect to intended analytical uses would substantially enhance transparency and interpretability [218,219,220]. Framing georeferencing outcomes as purpose-specific and context-limited spatial products, rather than universally transferable datasets, would strengthen comparability, reuse potential, and the long-term scientific value of historical imagery.

4.4. Emerging Technologies and Future Directions in Historical Imagery Referencing

Recent advances in computer vision, deep learning and automated feature matching have substantially expanded the methodological landscape for historical imagery georeferencing, offering new possibilities for addressing long-standing limitations associated with sparse control information, uncertain acquisition geometry and heterogeneous image characteristics [74,108]. In particular, modern computer vision approaches have improved the automation of key steps in the georeferencing workflow, including feature detection, feature description, correspondence matching, outlier filtering, and relative geometric reconstruction. Compared with traditional hand-crafted operators, learning-based methods can extract more robust and transferable image representations under varying radiometric conditions, local distortions, and partial scene changes, which are common challenges in historical aerial photographs and declassified satellite imagery. Deep learning–based matching frameworks have been especially influential in this context. By learning local descriptors and correspondence relationships directly from image data, these methods have shown greater tolerance to contrast variation, image noise, film degradation, and geometric inconsistency than conventional feature-based pipelines [56,74]. A representative example is SuperGlue [221], a graph neural network–based matcher that learns to establish sparse correspondences between keypoints across image pairs by jointly reasoning about visual appearance and spatial relationships. Applied to historical imagery georeferencing, SuperGlue and related frameworks have demonstrated the capacity to automatically identify spatially consistent candidate matches across temporally and radiometrically dissimilar image pairs [56]. These candidate matches can support subsequent tie-point or GCP selection while reducing manual intervention. Their potential value is particularly evident for historical imagery, where differences in illumination, sensor characteristics, image texture, and landscape appearance often reduce the reliability of conventional matching approaches. In addition, recent advances in automated feature matching and correspondence verification have improved the ability to establish well-distributed and reliable tie points across temporally separated or visually heterogeneous images, thereby increasing the feasibility of partially or highly automated georeferencing workflows in some historical settings. In parallel, increasingly automated Structure-from-Motion (SfM) pipelines have lowered the technical barrier for reconstructing relative camera geometry from collections of overlapping imagery, enabling internally consistent image block reconstruction even when explicit sensor metadata are unavailable [72,84,108]. These approaches can provide an effective basis for orthorectification or downstream spatial processing when image overlap is sufficient. Taken together, developments in computer vision, learned matching, and automated reconstruction suggest that higher levels of automation may be achievable for some categories of historical imagery, particularly in cases involving relatively stable landscapes, adequate image overlap, and sufficient image quality [53,56,57].

Despite these advances, the applicability of highly automated approaches remains constrained by several structural characteristics of historical imagery [40,128]. Many learning-based feature matching algorithms rely on training data that assume some degree of radiometric consistency or structural similarity between training and target images. Such assumptions are frequently violated in historical datasets, which may involve film degradation, non-standard acquisition geometries, and landscapes that have changed substantially over time [32,74]. Likewise, while automated matching can improve tie-point generation, false correspondences and spatial bias may still occur when repetitive textures, weak image structure, or major land-cover transformations reduce the distinctiveness of image features. Similarly, SfM-based reconstruction primarily recovers relative camera geometry and image structure. The resulting models are intrinsically scale- and reference-ambiguous, requiring additional constraints such as ground control points, reference DEMs, or ancillary spatial data to establish absolute georeferencing [51,108]. These limitations indicate that automated pipelines can streamline parts of the processing chain but do not eliminate the need for external geometric constraints or careful interpretation of the geometric relationship between image coordinates and ground coordinates in historical imagery referencing [3,52].

Future research is therefore likely to focus on approaches that improve the integration of automated methods with incomplete or uncertain historical information [40,128]. One promising direction involves AI-assisted reconstruction of missing metadata, such as approximate acquisition geometry, camera parameters, or flight trajectories inferred from image content and archival records [41,51]. Learning-based estimation of approximate camera models or acquisition configurations may help reduce the degree of geometric uncertainty in poorly documented image collections [74,222]. Another important direction is the design of matching and reconstruction frameworks that are explicitly adapted to temporal heterogeneity, scene transformation, and uncertain geometric priors in historical imagery. In addition, the emergence of large-scale vision foundation models and self-supervised representation learning offers further prospects for cross-domain image matching in data-scarce historical settings, where the availability of annotated training examples is inherently limited [223,224]. Uncertainty quantification in automated correspondence estimation also represents an underexplored but practically important research frontier, as explicit confidence measures could guide selective manual intervention and improve the reliability of georeferencing workflows operating across heterogeneous historical archives. More broadly, methodological development should prioritize hybrid strategies that combine computer vision and deep learning with external spatial constraints and domain knowledge, rather than assuming that full automation is universally attainable. In particular, such strategies should explicitly account for incomplete acquisition information and uncertain imaging geometry when reconstructing spatial relationships from historical imagery. Such advances would enable emerging technologies to extend the analytical value of historical image archives while remaining compatible with the data limitations inherent to historical remote sensing.

5. Conclusions

This systematic review of 205 peer-reviewed studies documents the methodological state of historical imagery georeferencing across a span of more than two decades of published research. The evidence points to three structural patterns that collectively characterize the field. First, georeferencing workflows remain predominantly manual or semi-automatic: semi-automatic procedures account for 112 studies and manual workflows for 89, while fully automatic pipelines appear in only four studies, all of which are restricted to declassified satellite imagery. Notably, the share of manual workflows rebounded to approximately 45% in the 2021–2025 interval, indicating that operator involvement remains common when georeferencing imagery with incomplete metadata or variable image quality. Second, ground control points constitute the primary geometric constraint in 89% of reviewed workflows, whether used alone (62 studies) or in combination with digital elevation models and sensor metadata (71 studies). GCP-free configurations remain rare and are largely confined to specific imagery types or well-documented acquisition contexts. Third, accuracy reporting practices are heterogeneous and often based on control-point residuals rather than independent checkpoints, limiting the comparability of reported positional errors across studies and reducing the capacity for cumulative methodological progress.

These patterns reflect structural properties of the historical georeferencing problem rather than gaps in available technology. Historical imagery was acquired under conditions such as non-metric film cameras, undocumented platform trajectories, absent calibration records, and landscapes substantially transformed since acquisition, which render the metadata assumptions underlying modern automated workflows inapplicable. The absence of a standardized georeferencing solution across the reviewed literature is therefore an expected consequence of irreducible data heterogeneity, not a correctable fragmentation. Geographic analysis further reveals that the evidence base itself is unevenly distributed: case studies concentrate in Europe and North America, while tropical regions, sub-Saharan Africa, and much of South America remain critically underrepresented despite their scientific value for establishing pre-satellite environmental baselines. This geographic skew limits the representativeness of the current evidence base and highlights the need for broader geographic coverage in future research.

These three structural patterns, namely workflow automation levels, GCP dependency, and accuracy reporting heterogeneity, directly inform the following methodological recommendations for researchers working with historical imagery datasets. First, workflow selection should be guided by data conditions, especially metadata completeness, the availability of stable reference features, terrain effects, and the degree of landscape change, rather than by software availability alone. Second, transformation models should be chosen on the basis of plausible imaging geometry, and higher-order local fitting approaches should be applied cautiously and explicitly justified. Third, accuracy assessment should rely on independent checkpoints rather than control-point residuals alone. Reported error metrics should encompass not only global summary statistics but also spatial error distributions. Furthermore, explicit fitness-for-purpose statements referenced to the analytical objectives of each study should be included. Together, these three recommendations constitute a practical baseline for improving methodological transparency and cross-study comparability. On the methodological side, continued progress will also depend on automated inference of missing acquisition parameters from image content and archival context. Further advances require uncertainty characterization that captures spatial error correlation rather than global average metrics. The development of GCP-free constraint strategies suited to regions where stable reference features are scarce or rapidly changing also remains a priority. On the geographic side, capacity building in underrepresented regions, combined with targeted digitization and metadata recovery efforts, would expand the evidentiary basis for method development and application. Establishing shared validation standards that are sensitive to the case-specific nature of historical data processing represents a realistic near-term goal, and the evidence synthesized here provides a concrete empirical foundation for that effort.