Location: Crop Production Systems Research
Project Number: 6066-22000-081-001-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: May 1, 2020
End Date: Apr 30, 2025
1. To establish the benchmark of remote sensing for potato yield mapping using vegetation indices. 2. To develop machine learning schemes/algorithms to improve potato yield mapping based on high-resolution satellite and UAV remote sensing. 3. To integrate and create a new remote sensing tool for on-farm potato yield mapping.
1. Experiment setup and data collection: A multiple-dimension data-fusion prototype for potato yield assessment is established to represent soil, landscape, management practices, remote sensing, and yield. The data for management practices include previous crops, planting, irrigation, nutrients, disease/pest control, and tillage. Soil data consist of gridded soil analysis in this or previous growing season and soil data from the Soil Survey Geographic Database (SSURGO) (USDA NRCS, Washington, DC). The LiDAR elevation data at 1-m resolution will be downloaded from the website of MN TOPO, Minnesota Elevation Data Web Application at http://arcgis.dnr.state.mn.us/maps/mntopo/. The schedule of management practice dates will also be recorded. Potato growth status in two potato farms will be monitored through daily satellite-based imagery data from the Planet Company (San Francisco, CA) and UAV remote sensing images. The PlanetScope satellites provide daily data at 3.7 m resolution and four spectral bands (red, green, blue (RGB) and near infrared (NIR)). The Quantix Hybrid UAV system (AeroVironment, Simi Valley, CA) will be used to collect multispectral images at key growth stages across the growing season. To better select the most representative sites for spatial yield variability for ground sampling and validation, the conditional Latin hypercube sampling (cLHS) by integrating all features of environment, agronomy and remote sensing monitoring will be used to determine 50 ground-truth sampling sites for potato yield measurement in each field. During potato harvest, the global positioning system (GPS) data for the cover areas/harvester paths for each full bulk truckload will be collected, and all truckloads will be weighed. Random samples from the truckloads will be used to determine mud/tare weight to calibrate the potato yield data based on truckload. 2. Yield prediction tool development: We will explore support vector machine (SVM), random forest (RF), deep neural network (DNN) and deep learning-long short-term memory (LSTM) recurrent neural network (RNN) for their performance of potato yield prediction based on fused information from environmental, agronomic and remote sensing monitoring features. To better examine the contribution from temporal monitoring information of remote sensing, three scenarios will be adopted to investigate: 1) keep all monitoring data during five growth stages, 2) keep the monitoring data only during tuber bulking, and 3) accumulate whole season monitoring. Totally, nine models of yield prediction with the datasets will be trained and validated. All data processing and machine learning tool development will be conducted on the supercomputing clusters at the Minnesota Supercomputing Institute (MSI). We will use a python-based machine learning packages for tool development and testing. 3. Yield prediction tool evaluation: Evaluation of yield prediction tool in each individual field and across two fields will be conducted through in-field validation and cross validation between fields for their performance of yield prediction using the potato yield data manually sampled and determined in each field.