Location: Southeast Watershed Research
Project Number: 6048-13000-027-52-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Aug 21, 2017
End Date: Aug 20, 2021
This project will bring together university and USDA scientists to resolve questions underlying the characterization of regional-scale geospatial models of agricultural systems. Central to this research will be the question of “representativeness” of agricultural experimental plots and fields to regions being defined by the Long Term Agroecosystem Research (LTAR) network. Regionality may be defined by properties of geography, climate, ecological function and agricultural production and management, among others. Inherent in this research question are inquires of spatio-temporal patterns and correlations of factors including temperature, precipitation, physiography, landscape metrics, ecosystem services and agricultural production, operating at multiple scales within and across LTAR regions. While the scope of this research is constrained to the southeastern coastal plain, an intended outcome is to develop methodology that can be extrapolated to LTAR locations across the continent, thereby providing a systematic rationale for designating the boundaries representative of LTAR sites. Anticipated results from this research will provide a mechanism by which management scenarios can be linked with biophysical variables and processes at plot, field and regional scales within the Gulf Atlantic Coastal Plain LTAR, and whereby fluctuating, emergent agro-ecosystem services (and disservices) can be quantified and mapped. Geographic and mathematical theories of scale, spatial statistics, biogeography and coupled human and natural systems (i.e. CHANS), combined with advanced techniques in machine learning and remote sensing, will form the foundations for the methods of this research. The work will be accomplished by Ph.D. students in Statistics and Geography supervised by faculty members from both Departments at the University of Georgia, and scientists at the USDA-ARS. The data gathered for this project will have spatial, temporal, and explanatory relationships to discover, describe, and assess. Working together, the students will use geographic information science and technologies (GIS&T), coupled with advanced spatial statistics. Traditional methods of image classification will be employed, along with emerging image analyses methods such as Deep Learning, the modeling of high level abstractions from multiple non-linear transformations that take into account spectral, spatial and temporal components of multi- to hyper-spectral resolution (i.e., 4 to 200+ spectral bands) and medium- to very high-spatial resolution (i.e., pixel sizes of I km to 2 cm) imagery acquired at time intervals from monthly to hourly. This research will provide an opportunity for combining the statistical theory of empirical orthogonal functions, a common method for modeling spatio-temporal processes, with techniques such as a self-organizing map, from the subfield of machine learning, artificial neural networks.
Cooperator - Multi-resolution remote sensing imagery acquired from satellites, airborne platforms, and small unmanned aerial systems will be used to characterize landscapes over regions, watersheds, and agricultural sampling and experimental plots. Deep Learning and Big Data analytics will be applied to landscape-level geographic data over time, (annual, seasonal and customized drone flights) to derive information on spatio-temporal trends in and relationships among climate variables, land use/land cover, hydrography, vegetation health/diversity, natural- and human-induced disturbances, vegetation productivity and crop health. Terrain characterization incorporating existing national, state and local digital elevation models, Long Term Agroecosystem Research (LTAR), digital surface models, and multi-image matching Structure from Motion techniques will be combined with high spatial, temporal and spectral resolution drone imagery. Geographic information system technology will integrate these data to perform geospatial modeling and analysis in 2D, 3D and 4D (i.e., time) dimensions. Working closely with the Ph.D. student in Statistics and USDA-ARS scientists, nested models of risk assessment will be developed that examine past, current and future land management scenarios given time-appropriate data while considering data quality, uncertainty and completeness. Seymour- Spatio-temporal processes modeling will be conducted using empirical orthogonal functions that break the spatio-temporal observations into uncorrelated spatial and temporal components and then integrated with a self-organizing map (SOM) technique of machine learning. This technique decomposes a training data set into nodes, and then any new observations beyond the training data are expressed and studied as a weighted combination of the nodes. The mathematical/statistical difficulty with SOMs is that the nodes are not independent components, and little is known about their statistical significance. This project will have spatial, temporal, and explanatory relationships to discover, describe, and assess the potential of mashing together the statistical theory of EOFs with the machine learning technique of SOMs. ARS- Synthesis will characterize and quantify uncertainty and variability at different scales of analyses that consider scenarios of change in human-induced and biophysical drivers using a hierarchy of models spanning several levels of input data quality and density. Models may be calibrated by “back-casting” scenarios to evaluate decadal-scale changes, and evaluated by determining the efficiencies associated with model outcomes as the drivers change through time. Outputs of model results, synthesized according to a theoretical framework (e.g. ecosystem services, resilience, etc.), will evaluate risk applicable to agricultural policies focused on sustainability and resilience (e.g., How do we provide NRCS, FSA, Congress, etc. with information on our confidence that a given practice, incentive policy, or regulation will manifest as an ecosystem service versus disservice, and what would be the relative cost-benefit of those decisions?).