Location: Southeast Watershed Research
Project Number: 6048-13000-028-002-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Jun 1, 2022
End Date: May 31, 2027
Via cooperative research among Universtity of Georgia and USDA scientists, this project addresses questions underlying the characterization of regional-scale geospatial models of agricultural systems. These fundamental geospatial science questions are being addressed by the Long-Term Agroecosystem Research Network (LTAR) to determine the extent of our areas of inference for research studies, and how well our study sites represent broader conditions in the surrounding region. Inherent in this research are investigations of spatio-temporal patterns of factors including temperature, precipitation, physiography, landscape metrics, ecosystem services, and agricultural production, at multiple scales within and across regions. Regionality may be defined by properties of geography, climate, ecological function, agricultural production and management, etc. LTAR scientists have focused on the question of “representativeness” of agricultural experimental plots and fields to regions, a recognized knowledge gap in the LTAR Network. The methodologies developed in this research will enable extrapolation to LTAR locations across the continent, providing a systematic rationale for designating boundaries representative of LTAR sites and supporting LTAR Network management. This project will use geospatial artificial intelligence (GeoAI) methods to determine the spatial constraints of scientific predictions and extrapolate outward from ARS research study areas to places where direct measurements are sparse or missing. This project will develop models consisting of Artificial Intelligence workflows and Machine Learning (ML)/Deep Learning (DL) algorithms to extrapolate key crop production variables to broader extents. Initial work on this research has been confined to single computing workstations, working with a limited set of multispectral reflectance data from one cropping season. To operationalize these methods, the models need to be refined to ingest multiple ancillary and large datasets, and expanded temporal collections within a high-performance computing (HPC) environment such as the USDA-ARS SCINet clusters. This project will optimize and scale current workflows for use in a parallel processing environment. It will improve and refine our ML/DL models by adding additional fine-scale information measured by sensors borne by uncrewed aerial systems (UAS) on the microtopography and within-field soil and vegetation conditions. This new agreement expands on the geospatial modeling of dynamic agro-ecological systems to include Artificial Intelligence innovations for scaling field-based measurements to the broader region. Our results provide a mechanism for management scenarios to be linked with biophysical variables and processes at plot, field and regional scales, whereby fluctuating, emergent agro-ecosystem services (and disservices) can be quantified and mapped within the Gulf Atlantic Coastal Plain LTAR and beyond.
The approach to this project will use statistical models combined with methods in geospatial artificial intelligence (GeoAI) to advance agroecosystem monitoring and predictive modeling. The cooperator will focus on data collection using sensors mounted on uncrewed aerial systems (UAS) available at UGA and USDA. Field sensors include high-resolution cameras, real time kinematics (RTK GPS), multispectral sensors, Light Detection and Ranging (LiDAR), thermal and hyperspectral sensors. The cooperator will compare the microtopography of agricultural fields derived from UAS LiDAR data and photogrammetric techniques such as multi-image matching Structure from Motion (SfM) with high spatial, temporal and spectral resolution drone imagery. Geographic information system (GIS) technology will integrate these data and provide a platform for performing geospatial analysis in 2D, 3D and 4D (i.e., time) dimensions. Image and geospatial data-derived training and validation samples will be used in GeoAI algorithms of Machine Learning (e.g., Random Forest), Deep Learning (i.e., Deep Neural Networks) and Explainable AI for image data classification and predictive modeling for applications in precision farming, yield monitoring/predictions, risk assessment and analysis of ecosystem services and rural societal benefits. The cooperator will provide subject area expertise in spatial statistics, GeoAI, remote sensing and photogrammetry. Advanced and novel statistical methods will be developed to resolve questions about extrapolating uncertainty beyond the bounds of direct measurement to support ARS goals of extending research results to areas and conditions where no direct measurements exist. Statistical downscaling models, multimodal data analysis and convolutional neural networks (CNNs) will be used to study the spatial, temporal and explanatory relationships of information derived from in situ data, UAS-mounted sensors and satellite remote sensing. The project will advance GeoAI for characterizing agricultural monitoring plots, fields, watersheds and regions using available satellite data, along with data collected with our UAS-mounted high-resolution cameras, RTK GPS, multispectral sensors, LiDAR, thermal and hyperspectral sensors. We will derive spatio-temporal trends and relationships among variables of climate, land use/land cover (LULC), micro-terrain, hydrography, vegetation health/diversity, natural- and human-induced disturbances, vegetation productivity and crop health at multiple scales. Working closely with our UGA-USDA colleagues, image/geospatial data-derived inputs to GeoAI algorithms will result in: 1) classification of crop types and phenology, weeds and surrounding LULC; and 2) predictions for precision farming, yield forecasting, risk/damage assessment, ecosystem services and rural societal benefits. We will focus on two ARS agroecosystem indicators of interest, Above Ground Crop Biomass and Above Ground Crop Development Stage. The project will culminate in the completion of at least four manuscripts in Statistics and Geography peer-reviewed journals.