Submitted to: Meeting Abstract
Publication Type: Abstract Only
Publication Acceptance Date: 2/12/2023
Publication Date: N/A
Interpretive Summary: On-the-go sensing is an efficient way of collecting field data on soil and agronomic properties of interest. Data collected from multiple sensing platforms and their spatial-temporal properties can benefit input optimization through precision agriculture and conservation practices. We collected corn yield, grain quality, soil apparent electrical conductivity (ECa), and optical sensor data on-the-go and used them to predict soil nutrient levels at multiple depths across a 46.2 ha field in Temple, TX. A machine learning model-random forest was used for the prediction, and the relationship between nutrient levels and yield and grain quality was also reported.
Technical Abstract: On-the-go sensing is an efficient way of collecting exhaustive information on soil and agronomic properties of interest. Field data collected from multiple sensing platforms and their spatial-temporal properties help identify patterns in the data that could benefit input optimization through precision agriculture and conservation practices. We combined corn yield, grain quality (protein and starch), soil apparent electrical conductivity (ECa), and optical sensor data (Red and NIR) collected on-the-go with measured soil properties to infer and map the later at multiple depths across a 46.2 ha field in Temple, TX, and evaluated their inter-relationships. Yield data was collected with an AgLeader yield monitor, protein and starch content with a CropScan 3300h grain quality monitor, and ECa and optical data with a Veris 3210 instrument. Bulk ECa data from the instrument was inverted and mapped using InVERIS software to generate a quasi-3D ECa profile, and the inverted values were aggregated to three depth intervals (0-30, 30-60, and 60-90 cm). One hundred and fifteen sampling locations were identified in the field based on ECa, Red, IR and field topography (wetness index) data using conditioned latin hypercube sampling design, and core soil samples were collected using a Veris P4000 probe. Soil samples from 0-30, 30-60, and 60-90 cm were analyzed for texture (sand, silt and clay %), organic matter (%), pH, cation exchange capacity (meq/100g), and primary nutrients (N, P, K- mg/kg) and for base saturation (%). A machine learning based model – random forest- was employed to predict and map measured soil properties using ECa, wetness index, slope, and optical data as predictors on 80% measurements, and the model was evaluated on the remaining 20%. Uncertainty of prediction, variable importance (%), and the relationship between soil properties and nutrient levels with yield and grain quality was reported.