Machine Learning Based High-Throughput Phenotyping Framework for Crop Yield Prediction Using Unmanned Aircraft Systems
Abstract
Estimating crop yield is essential to ensure agricultural stability, economic viability, and global food security. Provided with accurate crop yield estimation before harvest, farmers, breeders, and agriculture researchers can perform crop evaluation, genotype selection, and maximize yield by timely intervention. Remote sensing is often used to provide information about important canopy state variables for crop yield estimation. However, until recently, a critical bottleneck in such research was the lack of high-throughput sensing technologies for effective and rapid evaluation of expressed phenotypes under field conditions for holistic data-driven decision making. Recent years have witnessed enormous growth in the application of unmanned aircraft systems (UAS) for precision agriculture. UAS has the potential to provide information on crops quantitatively and, above all, nondestructively. This dissertation aims at utilizing UAS data to develop a machine learning based high-throughput phenotyping framework for crop yield estimation. In this research, plant parameters such as canopy height (CH), canopy cover (CC), canopy volume (CV), normalized difference vegetation index (NDVI), and excessive greenness index (ExG) were extracted from fine spatial resolution UAS based RGB and multispectral images collected weekly throughout the growing season. Initially, a comparative study was conducted to compare two management practices in cotton: conventional tillage (CT) and no-tillage (NT). This initial study was designed to test the reliability of the UAS derived plant parameters, and results revealed a significant difference in cotton growth under CT and NT. Unlike manual measurements, which rely on limited samples, UAS technology provided the capability to exploit the entire population, which makes UAS derived data more robust and reliable. Additionally, an inter-comparison study was designed to compare CC derived from RGB and multispectral data over multiple flights during the growing season of the cotton crop. This study demonstrated that using a morphological closing operation after the thresholding significantly improved the RGB-based CC modeling. A CC model that uses a multispectral sensor is considered more stable and accurate in the literature (Roth and Streit, 2018; Xu et al., 2019). In contrast, the RGB-based CC model is unstable and fails to identify canopy pixels when cotton leaves change color after canopy maturation. The proposed RGB-based CC model provides an affordable alternative to the multispectral sensors that are more sensitive and expensive. After assessing the reliability of UAS derived canopy parameters, a novel machine learning framework was developed for cotton yield estimation using multi-temporal UAS data. The proposed machine learning model takes three types of crop features derived from UAS data to predict the yield. The three types of crop features are multi-temporal canopy features, nontemporal features (cotton boll count, boll size, boll volume), and irrigation status. The developed model provided a high coefficient of determination (R2 ~ 0.9). Additionally, redundant features were removed using correlation analysis, and the relative significance of each input feature was determined using sensitivity analysis. Finally, an experiment was performed to investigate how early the model can accurately predict yield. It was observed that even at 70 days after planting, the model predicted yield with reasonable accuracy (R2of 0.71 over test set).
Degree
Ph.D.
Advisors
Jung, Purdue University.
Subject Area
Agriculture|Artificial intelligence|Management
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.