Abstract
Horseradish (Armoracia rusticana) is one of Illinois’ most important specialty crops, with an annual farm-gate value of approximately $10 million. Weed management is a major challenge in its cultivation due to restrictive herbicide application and reliance on manual labor. A weed detection system is needed so precision agriculture machinery can help farmers better manage their fields. This study aims to develop a non-AI method of weed detection using image features. A YOLO model was used to prelabel images as horseradish crop or weed. Within these images, the centroid of each plant was extracted, and a 50-pixel region of interest (ROI) was created around it. For each ROI, texture metrics were extracted using a gray-level co-occurrence matrix, revealing a noticeable split between crops and weeds based on mean texture. Using these mean texture metrics, a cutoff method was developed based on the range and distribution of texture values within each image. This included a dynamic approach for images with a bimodal distribution and a static approach, using a percentage cutoff value derived from 30 random images, for normal distributions. The cutoff methods were evaluated for error percentage across 402 images, yielding a mean error rate of 8.3%. A mean false negative rate of 3.6% showed a low rate of misclassifying crops as weeds. Overall, the method demonstrates a non-AI approach for differentiating weeds from horseradish in field images. This differentiation method requires less hardware, such as GPUs, lowering cost and making the technology more accessible.
Keywords
Horseradish, Weed detection, Precision agriculture, Non-AI classification, Texture metrics
DOI
10.5703/1288284318725
Non-AI Texture-Based Automated Annotation for Crops and Weeds
Horseradish (Armoracia rusticana) is one of Illinois’ most important specialty crops, with an annual farm-gate value of approximately $10 million. Weed management is a major challenge in its cultivation due to restrictive herbicide application and reliance on manual labor. A weed detection system is needed so precision agriculture machinery can help farmers better manage their fields. This study aims to develop a non-AI method of weed detection using image features. A YOLO model was used to prelabel images as horseradish crop or weed. Within these images, the centroid of each plant was extracted, and a 50-pixel region of interest (ROI) was created around it. For each ROI, texture metrics were extracted using a gray-level co-occurrence matrix, revealing a noticeable split between crops and weeds based on mean texture. Using these mean texture metrics, a cutoff method was developed based on the range and distribution of texture values within each image. This included a dynamic approach for images with a bimodal distribution and a static approach, using a percentage cutoff value derived from 30 random images, for normal distributions. The cutoff methods were evaluated for error percentage across 402 images, yielding a mean error rate of 8.3%. A mean false negative rate of 3.6% showed a low rate of misclassifying crops as weeds. Overall, the method demonstrates a non-AI approach for differentiating weeds from horseradish in field images. This differentiation method requires less hardware, such as GPUs, lowering cost and making the technology more accessible.