Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Abstract

Automated assessments of residue cover based on imagery potentially reduce labor and human bias associated with in-field measurements. Our objective was to evaluate 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 (2018 data) and 1,400 used for testing (2019 data). Percent residue for each ROI image (ground truth) was determined by a bullseye grid method (n = 100; Lory et al. 2021). The ROI images were divided into the maximum number of 224 X 224 sub-images. Features were extracted from each sub-image using the VGGNet16 model, a convolutional neural network model. To reduce the number of features, we only used features from the five max-pooling layers and averaged them based on each kernel. Average values for each kernel were then averaged across all sub-images of the ROI resulting in a 1,472-feature dataset per ROI. Two feature selection strategies were then applied to the features: recursive feature elimination support vector machine classification (RFE-SVM) and forward regression feature selection (FRFS). Best locations outcomes were obtained with RFE-SVM (Table 1).

There was no apparent pattern of correlation among selected features and no overlap in outliers suggesting unique characteristics among the two selected feature sets. Results were superior to previous research based on the same data set using 70 manually extracted known features (Upadhyay et al., 2022). Results also compare favorably with other previous published research estimating of residue cover using high-resolution RGB images (r² between 0.75 and 0.90). Implementing a model derived from transfer learning is more computationally complex than implementing a model derived from known manually extracted features. Success of the transfer learning model documented there is an opportunity to improve upon the features developed to date by more conventional methods. Transfer learning through features extracted from VGGNet16 was a successful strategy for estimating residue cover from high-resolution RGB imagery.