Integration of MATLAB and Machine Learning to Accelerate Evaluation of Biological Activity in Agricultural Soils and Promote Soil Health Improvement Goals
- Authors
- Publication Date
- Aug 01, 2024
- Source
- University of Nebraska - Lincoln
- Keywords
- License
- Unknown
- External links
Abstract
Traditionally, assessments of soil biological activity have been confined to laboratory settings, creating a disconnect with practical in-field methods. To bridge this gap, cotton fabric degradation has been used to illustrate soil microbial activity under different management practices. While effective, these demonstrations are subjective and labor-intensive. Researchers have explored using image processing software like ImageJ and Adobe Photoshop to streamline this process. Although these tools accurately quantified fabric degradation under varying soil conditions, the methods remained labor-intensive and complex. Consequently, these methods were still not ideal for on-farm use by agricultural practitioners. To further address labor and complexity limitations, the experiment in Chapter 2 aimed to develop a more controlled and replicable method for fabric photography and a software program to automate image processing and analysis. This new approach utilized the Scikit-image Python package and MATLAB.Engine, leveraging Python scripting to access MATLAB's image-thresholding applications. The process involved segmenting images based on specific colors in the HSV (Hue, Saturation, and value) space to accurately identify degraded areas. This technique was initially tested on a dataset of 39 images, resulting in a promising coefficient of determination (R2) of 0.8, indicating good predictive performance compared to manual Photoshop methods. However, the MATLAB-based method faced challenges such as sensitivity to variations in photographic settings and the need for expensive equipment and software licenses, making it less accessible for field use. To overcome these issues, the study described in Chapter 3 then explored using machine learning models: MobileNet, VGG16, ResNet-50, U-Net, and YoloV8, on a broader dataset of 176 images representing more diverse conditions. This strategy aimed to improve model robustness against background and lighting variations. The best results were obtained with MobileNet, which demonstrated excellent predictive performance (R2 = 0.944) and significantly reduced prediction time. Despite its effectiveness, the model still showed sensitivity to abrupt changes in image properties, which affected its accuracy (R2 = 0.904). Continuous refinement of the model is suggested to maintain its efficacy, highlighting the potential of machine learning to enhance the practical monitoring of soil biological activity in agricultural settings. Advisor: Amy M. Schmidt
