The Farm Security Administration - Office of War Information archive is a collection of over 170,000 monochrome and color photographs, commissioned between 1935 and 1945 by the United States federal government. The collection originated in 1935 as a publicity campaign for the Resettlement Administration, one of President Franklin Roosevelt’s Second New Deal projects centered on aid to migrants from the drought stricken southwest. By the late 1930s, the collection was moved to the Farm Security Administration and had grown into a massive documentary photography project with aspirations to document the entire United States. Known as the Historic Division, this division of photographers led by Roy Stryker was finally transferred to the Office of War Information in 1942 to serve government interests in World War II. Offering a unique snapshot of the nation during the period, the archive serves as an important visual record for scholars and the public-at-large. Additionally, the photographers and administrators involved with the FSA-OWI collection have been the subject of extensive academic scholarship around how they shaped America’s documentary record and documentary photography [Hurley 1972] [Cohen 2009] [Stott 1973] [Mendelson 2003] [Mora 2006].

Following World War II, the collection was transferred to The Library of Congress for cataloging. The Library of Congress chose the collection to experiment with digitization and digitized the entire collection in the 1990s. With changes in technology and new digital standards, they are in the process of re-digitizing the collection to current archival standards.[1] Since the photographs were taken on behalf of the United States Government, access to and use of the collection is essentially free and open. Given that it is a public resource, the Prints and Photographs Division of the Library of Congress has hosted archival quality scans of every photograph in the collection on their public website. Extensive metadata accompany a majority of the photographs, capturing features such as dates, titles, locations, photographers’ names, and a unique identifying call number. Basic text search features are provided, but otherwise the website is presented as a digital reproduction of the physical archive with a limited set of digital-only features.

While the existing interface allows for browsing access to the archived material, we understood that more direct access to the data would be necessary in order to analyze the records computationally. Neither an API nor a single downloadable website snapshot was provided by the Library of Congress. Given the clean structure of the Library of Congress website and copyright status of the FSA-OWI collection, creating a local copy of the archive via a custom webcrawler proved relatively straightforward. The complete code and documentation for this process are publicly available. Prior to working with the downloaded data, we cleaned the data in order to standardize each field. For instance, dates were written in a number of styles (“Aug 1943”, “Aug. 1943”, “August 1943”, “Summer 1943”) and needed to be transformed on a case by case basis into a common format. Other fields, such as photographer names and captions, had similar issues. County, city, and state names were particularly messy and required using a web-based geocoding service.

Once organized in a single local database, the FSA-OWI collection could be viewed in aggregate. The ability to filter and tabulate by various parameters provided new insight into the collection's scope. For example, creating a table showing the number of images taken by each photographer showed how relatively few photographs some of the more well-known staff photographers took (Walker Evans was only the 20th most populated) and exposed the extensive number of contributing photographers (over 54 have at least 50 photographs in the collection, and over 133 were credited overall). Creating a table of photographs per state revealed that no fewer than 123 photographs (Rhode Island) were taken in each of the then extant 48 states. Viewing by country revealed hundreds of photographs from overseas, including North Africa and the Caribbean.

In addition to aggregate counts, having the archive in a local database also enabled the application of digital humanities methods to close read the collection, with a focus on spatial and computational analysis. Often, computational analysis is applied to study an archive at scale. Scholars such as Franco Moretti [Moretti 2005] and Matthew Jockers [Jockers 2013] have pioneered and refined this approach. However, computational analysis can also enable close reading, a fundamental aspect of humanities scholarship [Abello, Broadwell, and Tangherlini 2012]. This is important because often skepticism regarding the field of digital humanities centers around concerns that close reading is compromised through computational methods, a false dichotomy problematized and addressed by Michael Hacher [Hancher 2016]. The methods applied to the FSA-OWI archive as explored in this article show how computational methods can play a critical role in close analysis of sources and further humanities scholarship.

Our computational analysis led to three major components of Photogrammar. As with the aggregate tabulations, spatial representation (mapping) of the archive helped distill the general patterns of what, where, and when the photographers were documenting. Mapping also provides information local to individual entities in the collection, thus assisting scholars interested only in a close reading of those particular photographs. After realizing that mapping could assist in the study of the FSA-OWI collection at a variety of scales, we sifted through our local copy of the archival database and discovered several additional latent features in the photographs and metadata. For example, we were able to impute the historic hierarchical archival developed in the 1940s by Paul Vanderbilt and we were also able to reconstruct strips of 35mm film used by FSA-OWI photographers. These three methods for uncovering latent metadata in the FSA-OWI collection are described in detail in the remainder of this article, with a focus on the kinds of contextual depth they add to understanding individual records within the archive as well as in the aggregate.

Just over half of the photographs in the digital FSA-OWI collection are tagged with some level of geographic information. In some cases this was as detailed as the name of a particular street in a large city. The majority, however, were referenced at the county or city level. Using a geocoding service, along with some careful tweaks to adjust for changes in historical county borders, we tagged the photographs with latitude and longitude coordinates. Plotting these over a modern map obfuscates the routes staff photographers traveled in the 1930s and 1940s because many of the modern-day highways are part of the Interstate Highway System, which was constructed in the latter half of the 20th century. Instead, the photograph locations were mapped over a Vico Motor Oil Map from 1937 (Figure 1) . Beyond looking at the collection at scale, there is a wealth of information that can be learned and questions that can be posed about each photograph or sequence of photographs by putting them in geographical context.

For example, mapping all of Arthur Rothstein's photographs taken over the course of the summer of 1936 show that he was driving along major roads, primarily Route 10, between Washington State and Minneapolis-Saint Paul (Figure 2). The same map also highlights an outlier, as indicated by the pink dot near the town of Lark. Upon closer inspection, the photos are near the neighboring town of Carson, which is nine miles southwest of Lark. This observation is interesting because Rothstein needed to travel several hours out of his way in comparatively rural counties, something he did not appear to do elsewhere along his two-and-a-half month trek across Washington, Idaho, North Dakota, and Minnesota. As a result, one questions why he deviated from the main highway at this particular moment. Looking even closer, the photographs themselves begin to offer answers. The photographs near Carson are a sequence of five photographs of Mike Sullinger and his drought-ridden farm in July 1936. Here Rothstein finds a set of conditions to communicate the impact of the drought on farming and an individual farmer. By mapping a sizable set of the archive to discern Rothstein's route across the Northwestern US, we have been provided useful contextual information regarding a small sequence of five photographs.

Jack Delano also offers another example for the impact of mapping on close reading images. A significantly longer sequence of photographs is shown below, when Jack Delano traveled between Chicago, Illinois and Los Angeles by rail (Figure 3). Since this trip was completed entirely within the month of March of 1943, and dates in the collection were only recorded by month and year, it was not possible to order Delano's photographs from start to finish along this route without using a map. In turn, mapping allows the viewer to key the photographs to the map and see the order in which the photographs were taken. Such insight allow for a closer analysis of Delano's shooting process.

Paul Vanderbilt joined the FSA office in 1942. Under the direction of Roy Stryker, Vanderbilt developed two pioneering archival systems for cataloging the photographic prints produced by the photographic division: a classification system and a lot system.

For the classification system, he first split approximately 88,000 photographs into five geographic regions based on where they were shot. They were then assigned to a three-tier hierarchy based on the photograph's themes or subjects, a process today that one might refer to as tagging. The first level included 12 main subjects: The Land, Town and Cities, People, Homes and Living Conditions, Transports, Work, War, Medicine and Health, Religion, Intellectual and Creative Activity, Social and Personal Activity. The table of images below shows samples from each of the categories (Figure 4). After being assigned to a main subject, each photograph was then assigned to one of 50 sub-headings and then again assigned to one of approximately 1300 sub-sub-headings. The resulting system persists today as the method for organizing the photographic prints at the Library of Congress. Vanderbilt's hierarchical scheme and methodology for classifying photographs has been the subject of research in its own right [Trachtenberg 1988].

The Vanderbilt classification system tags are not overtly discernible from the metadata in the digitized archive. However, it is possible to reconstruct them using the available metadata fields. The unassuming field “Other Number” contains cryptic codes such as “D 915”, which represent the Vanderbilt categorization of the record. The only known mapping between these codes and the categories is a printed manual from 1981 [United States 1981]. Therefore, populating the local database of the archive with the classification system required obtaining this manual, creating a digital scan of the codes, running an OCR detector over the scans, and joining the resulting codes to the “Other Number” field.

Digitally reconstructing the categorization scheme greatly simplified the process of analyzing Vanderbilt's cataloging project. It was previously possible only when visiting the physical FSA-OWI archive and required substantial effort to look at several categories at once. Figure 5 shows an example of the kind of visualizations that can now be quickly constructed. All of the photos shown here were placed into the “People As Such” top level and “Groups and Individuals” second level categories. The first row was placed into the tertiary category “Men in Their Prime”, the second as “Middle Aged Men”, and the third as “Old Men”. Displayed as such, these sub-categories illustrate that the process of tagging photographs reflects subjective differences of interpretation. The caption of the second to last photograph on the first row, for example, indicates that this is in fact a photograph of Mrs. Janet P. Murray.

While the classification system offers one way to tag and read the archive, the lot system offers a different point of entry based on another piece of latent data, which resides in the MARC field “Notes”. Lots, also a creation of Paul Vanderbilt, were an organizing mechanism to provide a different way of accessing the collection. In this case, the photographs were put into a lot primarily based on a shooting assignment. Often a photographer would visit the same place more than once. For example, Lot 4 includes 124 photographs by Russell Lee in Salinas, California taken at multiple times between December 1941 and December 1942.

Surfacing and making visible the Vanderbilt system has given users digital access to the collection through the archival system developed in the 1940s. Vanderbilt's two systems are not only an exciting way to close read a subset of the archive but serve as an object of study themselves. The very categories that make up the schema reflect the type of knowledge the federal government believed it was collecting. Naming the subject of a photograph is a complicated negotiation that is often as much a reflection of the categorizer and time period as the image itself. As a result, the Vanderbilt system itself offers a way to close read the archive.

The large number of printed photographs of the FSA-OWI collection displayed in long filing cabinets at the Library of Congress is actually only a limited subset of the images in the digital archive. The Library of Congress also retained a larger set of the negatives; it was these that were carefully digitized and placed online.[2] While some negatives are from various large format cameras, the majority come from long rolls of 35mm film. When 35mm film was processed in Washington D.C., the FSA-OWI staff would cut the film into strips of 4-6 frames in order to facilitate the process of curating and printing the images. Conveniently, this film structure is hidden as the suffix in the “Reproduction Number” field associated with each image in the digital collection. For instance, the negative associated with the number LC-USF33-016166-M2 is on a single strip of 35mm file between LC-USF33-016166-M1 and LC-USF33-016166-M3.

From our local copy of the archive, it was a straightforward process to reconstruct the strips of film from these Reproduction numbers and to expose additional latent information. For example, take the reconstructed strip of film in Figure 6. In this particular strip, the middle frame is attributed to Carl Mydans, from February 1936, with the caption “Elbert Peets, Washington, D.C., Principal Town Planner, Milwaukee project, Suburban Resettlement Division”. There is no assigned photographer name, date, location, or caption to the four remaining images. Using just the latent data from the reproduction number and assuming that film was not passed between photographers, we can confidently attribute all five images to Carl Mydans. It is also reasonable to make an educated guess that all five photographs were also taken around February 1936, even though it is physically possible that the first two were taken before and the last were taken after this date. These attributes can be filled in algorithmically, allowing for scholars studying Carl Mydans in 1936 to locate all five photographs, rather than just the middle photograph. Because most of the archive's metadata comes from the physical prints rather than the negatives, over 65% of the 35mm images in the digitized collection contain no photographer names, locations, or dates. Using the Reproduction Numbers this count can be reduced by over half to only 31% without any photographer information.

In addition to filling in missing metadata fields, there is even more that can be gained from arranging the negatives as film strips. By actually investigating the photos in Figure 6, it becomes apparent that the first three images are in fact images of the same person. The one extant caption provides valuable information; this individual is Elbert Peet and he holds the position of Principal Town Planner. We also are able to put together the sequence of photographs and thus gain insight into the photographer's ways of seeing. Finally, having the sequence together allows for insight into what Roy Striker, who typically decided which photographs to print, was looking to find in the images, because we know that he choose to print the last image of Elbert Peets rather than the the first two. It should be noted that it would have been difficult to make this connection based purely on digital methods and requires closely looking at the records, yet at the same time, it would be incredibly difficult without digital methods.

Figure 7 shows an example of going further by putting individual strips of film together in order to start reconstructing an entire roll of film. Using computational methods, we were able to pull out the first two rows of film as possibly being sequential as they both contained photographs by John Vachon taken in July, 1941 in Kenosha County, Wisconsin. The second and third rows similarly contained images taken by Vachon that month from Chicago, Illinois. Once these three film strips were pulled out from the thousands of possible film strips, a close reading of the images revealed that the cheese shop in the top right corner is the same shop shown in the left middle pane, and the man sitting on the sidewalk at the end of the middle row is the same person shown at the start of the third row. As with the Mydans set, this reconstruction of the film adds contextual information regarding decisions by both the photographer and the archivists. The order of the photographs reveals additional latent information; it is now clear that Vachon traveled from Wisconsin and into Chicago, which was not discernible otherwise because the metadata for the photographs includes only month and year. John Vachon went on to take hundreds of photos in Chicago, depicting it primarily as a prosperous city. It is very interesting to be able to put the store fronts and bustling movie theaters in those photographs into perspective after knowing that the first scene he choose to capture of the city was a man slouched over, sitting on his suitcase by the side of the road. One can begin to ask questions about Vachon’s juxtaposition of scarcity and abundance in an American city.

The three examples addressed show where large-scale digital methods have uncovered contextual latent metadata and how this new information can assist scholars in close reading a subset of records. Excited about the possibilities of these methods, we felt that it important to provide a user-friendly platform for exploring the collection to a broader set of publics. A digital, public humanities project, Photogrammar provides such a space. It was built in consideration of other popular enhanced digital archives, including Library of Congress’s American Memory, University of Richmond’s Visualizing Emancipation, and HistoryPin. We have borrowed insights from these projects in an attempt to meet the needs of casual users broadly interested in 20th Century American history and photography alongside research scholars who may have been studying this specific collection for decades.

In order to build the Photogrammar website, an interdisciplinary team, including the present authors and Ken Panko, was funded by an NEH Digital Start-Up Grant. The size and scope of the collection made it extremely difficult to navigate without deep subject matter expertise across a number of disciplines. For example, it would have been difficult to discern the latent data sitting within the archive without specialists in data analysis alongside experts with an intimate understanding of the FSA-OWI collection. Drawing on disparate and complementary expertise, our interdisciplinary core team included a GIS specialist, instructional technologist, statistician and two historians.

In order to manage expectations and goals, we created a series of phases. After three years of work, including several week-long trips to the Library of Congress, we publicly launched the first iteration of the site in May 2014. This phase included launching Photogrammar with advanced search features, historical interactive maps, and a dynamic visualization of Vanderbilt's historic cataloging system. Additional features are in process in order to interact with an increasingly wider set of publics, including journalists, artists and K-12 students, and to further assist in active scholarship surrounding then collection and the historical era it documents. Made possible with support from ACLS, our current phase integrates interviews by the FSA-OWI staff photographers from the Smithsonian Archives of American Art, digitizes texts from the Federal Writer's project, and builds a public interface for the film strips. As we move forward, collaboration has served as a guiding principle and allowed the project to flourish in exciting directions.