Crop assessment and monitoring using optical sensors

Abstract

Crop assessment and monitoring is important to crop management both at crop production level and research plot level, such as high-throughput phenotyping in breeding programs. Optical sensors based agricultural applications have been around for decades and have soared over the past ten years because of the potential of some new technologies to be low-cost, accessible, and high resolution for crop remote sensing which can help to improve crop management to maintain producers’ income and diminish environmental degradation. The overall objective of this study was to develop methods and compare the different optical sensors in crop assessment and monitoring at different scales and perspectives. At crop production level, we reviewed the current status of different optical sensors used in precision crop production including satellite-based, manned aerial vehicle (MAV)-based, unmanned aircraft system (UAS)-based, and vehicle-based active or passive optical sensors. These types of sensors were compared thoroughly on their specification, data collection efficiency, data availability, applications and limitation, economics, and adoption. At research plot level, four winter wheat experiments were conducted to compare three optical sensors (a Canon T4i® modified color infrared (CIR) camera, a MicaSense RedEdge® multispectral imager and a Holland Scientific® RapidScan CS-45® hand-held active optical sensor (AOS)) based high-throughput phenotyping for in-season biomass estimation, canopy estimation, and grain yield prediction in winter wheat across eleven Feekes stages from 3 through 11.3. The results showed that the vegetation indices (VIs) derived from the Canon T4i CIR camera and the RedEdge multispectral camera were highly correlated and can equally estimate winter wheat in-season biomass between Feekes 3 and 11.1 with the optimum point at booting stage and can predict grain yield as early as Feekes 7. Compared to passive sensors, the RapidScan AOS was less powerful and less temporally stable for biomass estimation and yield prediction. Precise canopy height maps were generated from a CMOS sensor camera and a multispectral imager although the accuracy could still be improved. Besides, an image processing workflow and a radiometric calibration method were developed for UAS based imagery data as bi-products in this project. At temporal dimension, a wheat phenology model based on weather data and field contextual information was developed to predict the starting date of three key growth stages (Feekes 4, 7, and 9), which are critical for N management. The model could be applied to new data within the state of Kansas to optimize the date for optical sensor (such as UAS) data collection and save random or unnecessary field trips. Sensor data collected at these stages could then be plugged into pre-built biomass estimation models (mentioned in the last paragraph) to estimate the productivity variability within 20% relative error.