Page 86 - GIS for Science, Volume 3 Preview

REACHING FOR THE CLOUD: THE BIG WHY
National forest inventories (NFIs), such as the ones conducted by FIA, provide a gold standard for the estimation of forest status and attributes useful for planning and decision-making at multiple scales. While FIA’s collection of content is well managed, these resources are not suitable for providing information and knowledge to a wide spectrum of users. The FIA estimation tools require an understanding of relational databases to create relevant estimates. Additionally, users increasingly require information at finer spatial and temporal scales for which NFIs were not designed.
The internet has demonstrated the public’s growing competency with maps, and FIA has a long history of exploring map-based approaches to sharing information and knowledge. Modeling frameworks that integrate FIA plot data with ancillary geospatial data, such as satellite imagery, can also play an important role in FIA estimation. Artificial intelligence and machine learning (AI/ML) algorithms show promise in accounting for complex relationships between forest attributes and ancillary data. In particular, deep-learning techniques could provide important advances in FIA estimation procedures for small areas and change analyses.
To meet the user’s needs in timely ways, FIA must increase dramatically the pace and scale of computation. It currently takes months to convert field observations into information and estimates. We need to publish future map-based content—likely dozens if not hundreds of geospatial data layers —simultaneously with the release of tables to the public. These maps will include current estimates accompanied by geospatial information that addresses other agency priorities, such as the 1990 baseline for carbon accounting.
Given the focus on delivering map-based content, our ideal computing environment is tuned to facilitate parallel processing of geospatial data through efficient scripting and implementation of AI/ML techniques. This capability will likely require a transformation of our business models away from a traditional focus on desktop modeling and estimation into new approaches (and languages) that efficiently use the potential of the cloud to implement deep-learning applications.
The maps derived from the integration of FIA data and ancillary predictors are peer- reviewed, transparent, and shareable modeling approaches, and the models are constructed to simultaneously protect plot security and provide reliable estimates of various forest attributes across a range of spatial scales. These ancillary predictors are truly massive datasets that are proven to improve the precision of estimates based on FIA plot data. For example, incorporating structural information related to the height of the forest canopy adds value to these models, but this work cannot be accomplished in desktop environments.
At the same time, the resulting data and information are hosted and maintained in cost-effective ways that empower FIA to create the broadest range of products imagined. The goal is to create and host the data in a manner that reduces data migration and related concerns about keeping content current and error-free. This goal is particularly important because it relates to the various systems of record across the Forest Service in multiple disciplines. Central to the goal is identifying appropriate points of simple yet consistent integration into enterprise information systems to maximize the value of the FIA content. The result is increased capacity and understanding about how to maximize FIA’s use of geospatial technology and data. Maps of individual forest attributes can then be combined easily with other content to provide integrative assessments (developed by the agency and its partners) relevant to resource managers, policy makers, and the public that holds them accountable.
Put simply, the future of FIA’s estimation, reporting and decision-support requires access to new, computing-intensive resources to fully integrate Forest Service data with the wide varieties of other content relevant to create authoritative models and maps of forest attributes at the scale of continents. This collaborative work capitalizes on recent innovations in GIS and remote sensing technologies to integrate a growing archive of satellite-based Earth observation platforms such as Landsat—the longest continuous record of our changing Earth—and to enable efficient processing and modeling of these massive archives at the national scale, to geo-enable one of the largest, most detailed archives of forest inventory information in the world.
An early experiment in parallel processing for image analysis. These devices were originally for Android-based video streaming, designed to be plugged into the HDMI port of a television set. Technologists at the Forest Service installed Linux and the
R statistical package on each of the devices and set them up on a detached WiFi network where they shared a USB drive for storage. To maximize performance and run them at the fastest clock speed, they placed a fan beneath the units to cool down the CPUs. While the system worked, it wasn’t a very scalable solution and was ultimately deployed using ArcGIS Enterprise and raster analytics.
74 GIS for Science