Estimation of crop residue cover in high-resolution RGB images using features from a pre-trained convolutional neural network
Abstract
Plant residue on the soil surface increases the sustainability food and fiber production in agricultural systems. Automated assessments of residue cover based on imagery has the potential to reduce labor and human bias associated with in-field measurements. Our objective was to evaluate the capacity of a transfer learning strategy to improve estimates of residue cover derived from high-resolution RGB images. The imagery for the project was collected from 88 locations in 40 row crop fields in five Missouri counties between mid-April and early July in 2018 and 2019. At each field location, 50 contiguous 0.3 m x 0.2 m region of interest (ROI) images (ground sampling distance of 0.014 cm pixel-1) were extracted from imagery resulting in a dataset of 4,400 ROI images; 3,000 used for cross validation and training (data collected in 2018) and 1,400 used for testing (data collected in 2019). The percent residue for each ROI image (ground truth) was determined by a bullseye grid method (n = 100). Features were extracted from ROI images using the VGGNet16 model, a convolutional neural network model. To reduce feature numbers, we averaged the features based on each kernel resulting 1,472 feature dataset per ROI. After the extraction, we compared three feature selection strategies: recursive feature elimination support vector machine classification (RFE-SVM), sequential forward feature selection classification (SFFS-SVM) and forward regression feature selection (FRFS). Best locations outcomes were obtained with RFE-SVM (r2 = 0.93, MAE = 4.9, with three outliers) and FRFS (r2 = 0.94, MAE = 5.2, with two outliers). The three models had no apparent pattern of correlation among selected features and limited overlap in outliers suggesting unique characteristics among the three selected feature sets. These results were superior to previous research based the same data set using 70 manually extracted known features. This suggested that transfer learning through features extracted from VGGNet16 pre-trained on ImageNet was a successful strategy for estimating residue cover. This research also confirmed the utility of high-resolution RGB imagery to quantify residue cover in agricultural systems.
Degree
M.S.