Combining Machine Learning and Handcrafted Image Processing to Identify Common Weeds in Cotton Fields

Weed management is a critical aspect of agriculture, with weed infestations posing significant challenges to crop yields. Conventional manual methods for weed identification have limitations, emphasizing the need for innovative solutions. This research focuses on the integration of machine learning and handcrafted image processing to identify common weeds in cotton fields, with a specific emphasis on Southern India. The targeted weed species include Palmer Amaranth (Amaranthus palmeri), Pigweed (Amaranthus species), Johnsongrass (Sorghum halepense), and Yellow Nutsedge (Cyperus esculentus), which are prevalent in cotton fields. Through the use of an RGB camera and a developed ImageJ plugin for shape feature extraction, 39 shape features were extracted for each weed species. Three machine learning models, k-Nearest Neighbor (kNN), Random Forest (RF), and Support Vector Machine (SVM), were employed, and the results were compelling. RF exhibited the highest accuracy at 96.1%, demonstrating remarkable precision, recall, and F1- Score values, particularly for Palmer Amaranth. The practical implications of this methodology extend to farmers, producers, and other stakeholders in cotton production. The open-source nature of the tools used makes this approach accessible, cost-effective, and user-friendly. It offers a valuable resource for precise weed identification, timely intervention, resource optimization, and ultimately, improved crop yields. This research underscores the potential for advanced machine learning models in weed management, offering a practical solution for enhancing agriculture's productivity and sustainability.