Scaling agricultural field measurements to drone and satellite imagery using deep learning to assess the regional boundaries of the Long-Term Agroecosystem Research network

Authors: STONE, AUSTIN - University Of Georgia; POWELL, JOSEPH - University Of Georgia; Coffin, Alisa; MADDEN, MARGUERITE - University Of Georgia; SEYMOUR, LYNNE - University Of Georgia; AHN, JEONGYOUN - University Of Georgia

Submitted to: Interagency Conference on Research in the Watersheds
Publication Type: Proceedings
Publication Acceptance Date: 6/15/2021
Publication Date: N/A
Citation: N/A

Interpretive Summary:

Technical Abstract: Studying long-term agricultural trends is crucial for understanding the future of food security in response to increasing global population, climatic variations, and economic demand. The United States Department of Agriculture, Agricultural Research Service (USDA-ARS) currently operates the Long-Term Agroecosystem Research (LTAR) Network to research the sustainable intensification of agriculture in the conterminous United States. The network includes the Gulf Atlantic Coastal Plain (GACP) LTAR site, a region spanning across South Georgia and into both Florida and South Carolina within the Southeastern Plains. The GACP is managed by the ARS Southeast Watershed Research Laboratory in Tifton, GA, where scientists have been observing cropping, hydrologic and climatic systems in the 334 km2 Little River Experimental Watershed since 1968. This study aims to integrate remotely sensed data acquired by satellites, aircraft, and unmanned aerial systems (UAS) with field-based measurements to better understand how well the greater coastal plain area is represented by measurements within the GACP LTAR boundary. Using this approach, we will use remotely sensed data and climate variables to derive vegetation indices, crop cover classifications and indicators of ecosystem services at field, landscape and regional scales. We will characterize landscape structure and processes operating within the GACP with the goal of identifying appropriate scaling methods for the extrapolation of agricultural research results, using statistical methods to identify, validate and guide the improvement of scaling models. The coastal plain region will be evaluated in terms of its representativeness of GACP LTAR research results so as to identify realistic boundaries for the inference of results. To accomplish this, we will explore statistical and deep learning methods, such as convolutional neural networks and self-organizing maps, resolving multiple variables across various scales to produce a regional classification. In turn, the deep learning algorithms can advance efforts to define agroecoregions for the continental network of LTAR sites.