Performing Department
(N/A)
Non Technical Summary
The key goal of this project, codenamed QUENCH, is to estimate soil moisture content (SMC) at the root zone level for dryland farming. These SMC maps will be produced at 30m spatial resolution every 2-3 days. We stratify the target root zone for SMCs based on soil depth increments of 5, 10, 20, 50, and 100 cm. These choices are informed, and constrained, by the availability of in situ observations at these depths and are sufficient to capture fine-root productivity dynamics critical for plant growth. These field level SMC maps can be used by growers to inform their decision-making process throughout the crop season. Our study region is informed by data collated by the USDA identifying winter wheat areas that are in drought prone areas; we further restrict our study area to those regions where winter wheat is the main crop. Over 83% of the area where winter wheat is located experiences some form of drought, including exceptional, extreme, severe, or moderate conditions. Our target area encompasses the states of Colorado, Kansas, Nebraska, Oklahoma, and Texas. The SMC-based decision-making that the effort will support includes, (1) when to plant seeds, (2) what crops to grow, and also (3) when to harvest early and grow some other crop. In several areas, such as the High Plains Aquifer where groundwater sources have been severely depleted, irrigated farms are expected to transition to dryland farming.
Animal Health Component
40%
Research Effort Categories
Basic
40%
Applied
40%
Developmental
20%
Goals / Objectives
Arid and semiarid regions represent approximately 25% of CONUS and a significant fraction of arable land around the world. Considering the recent expansions of drought-impacted areas and the increased shortage of groundwater, exclusively rainfed arable land is expected to increase with transitions from irrigated to dryland agriculture.The key goal of the proposed effort, named QUENCH, is to estimate SMC by combining the complementary strengths of in-situ SMC sensors, remote sensing, scientific models, and novel deep learning techniques to produce SMC maps at the root zone level for dryland farms. These SMC maps will be produced at 30 m spatial resolution every 2-3 days.While our methodology does not preclude applicability to the entire CONUS (Contiguous United States), our study region focuses on areas where USDA has identified winter wheat as the primary crop.Over 83% of the area where winter wheat is located experiences some form of drought, including exceptional, extreme, severe, or moderate conditions. Our target area encompasses the states of Colorado, Kansas, Nebraska, Oklahoma, and Texas. Finally, we do not consider fields when they are snow covered (typically between December through mid-March).We identify four interconnected objectives to accomplish our main goal:Objective 1: Design models that capture variability and interactions across soils, topographical and meteorological variables, and remote/in-situ sensing to estimate SMC.Objective 2: Infuse domain knowledge in the model training process while extracting value from the limited number of in-situ observations.Objective 3: Model refinements at scale to cope with spatial, topographical, and meteorological variability while ensuring effective resource utilizations.Objective 4: Designan intuitive cyberinfrastructure for scientists, growers, and other users including stakeholders to explore dimensions associated with SMC and decision-making.
Project Methods
We will use deep neural networks (DNN) to produce SMC maps. DNNs offer three key advantages. First, they are exceptionally well-suited to assimilating and extracting patterns from high-dimensional, multimodal data while being robust to outliers. This capability is aligned with the diverse data types we consider including digital elevation models (DEMs), soil types, meteorological information, and hyper/multispectral data. Second, once trained, the DNNs produce inferences at high throughput without parameter calibration. Finally, underpinned byrepresentational learning (automated extraction of higher order features from the data) and the generation of hierarchical representations to capture highly nonlinear interactions between the input feature space, DNNs offer the potential to scale to large geographical extents.