Agricultural crop residue cover estimation using image analysis and machine learning

Management of prior seasons crop residue is an important agronomic and sustainability consideration for row-crop agriculture. However, current methods of estimating residue coverage require human inference and, thus, are typically non-systematic in their on-farm implementations. This research presents a straightforward and efficient method of collecting labeled crop residue coverage imagery data using on-board sensors while performing tillage operations. The manual line-transect measurement method used for labelling is assessed for accuracy to understand its limitations, and the resulting images are analyzed for potential automatic estimation methods. An adaptive thresholding algorithm based on Otsu's method is then presented to estimate the Crop Residue Cover (CRC) levels for a given image. An evaluation metric, the baseline sigma, was introduced which was derived from the universal deviation value of $6.1\%$ obtained from the line-transect accuracy verification experiment. Results using this metric shows an accuracy of $85\%$ on the test dataset. The adaptive thresholding algorithm is also compared and shown to outperform other similar low-cost methods such as global thresholding and standard Otsu's method on the same test images. Also in this study, a deep learning-based approach to crop residue classification using the ResNet-18 model is presented as an alternative to the adaptive thresholding method. The model was fine-tuned on a crop residue dataset and achieved a 98% Top-1 accuracy during training and a 94% Top-1 accuracy on the test dataset. An evaluation metric, the delta ± 1 accuracy criterion, was introduced, which allows for a small tolerance in classification, accounting for cases where the model's prediction is close to the ground truth. By applying the delta ± 1 accuracy criterion, the model's performance increased to 98% accuracy on the test data. To demonstrate the model's applicability in real-world situations, tillage efficacy and uniformity were tested by segmenting a test image and analyzing the predictions. The results showed the model's effectiveness in classifying crop residues under varying field conditions and its potential for assessing the uniformity of tillage operations. The deep learning-based approach offers a valuable tool for improving crop residue management, addresses the limitations of traditional methods and provides a more accurate and efficient solution for crop residue monitoring and decision-making processes.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Agricultural & Biological Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Folorunsho, Samuel Oluwadare

Contributors dc:contributor

• Allen, Cody M
• Schwing, Alexander G
• Grift, Tony

• Crop Residue
• Image Processing
• Computer Vision
• Soil Conservation
• Adaptive Thresholding
• Machine Learning
• Residual Neural Network

Statement dc:rights

• Copyright 2023 Samuel Folorunsho

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/121188

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F121188