Enhanced Crop Weed Detection Using an Optimized Swin Transformer Architecture for Precision Agriculture

Mehendale, Nachiket Anil (2025) Enhanced Crop Weed Detection Using an Optimized Swin Transformer Architecture for Precision Agriculture. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (473kB) | Preview
Preview	PDF (Configuration Manual) Download (530kB) | Preview

Abstract

Under the framework of precision agriculture, machine learning technology enables better crop handling, increased agricultural output, and reduced harm to the environment (Liakos et al., 2018). A key use of this technology is to accurately tell the difference between crops and weeds on farms, so machines can remove weeds automatically and reduce the need for chemical weed killers (Mortensen et al., 2016). Current machine learning approaches for this task include Convolutional Neural Networks (CNNs), Vision Transformers, and Swin Transformers (Dosovitskiy et al., 2021). Among these, Swin Transformers work better for crop-weed detection because they process both detailed local features and broader image context in a hierarchical manner (Liu et al., 2021). However, Swin Transformers require significant computing power, which limits their use on farms with basic hardware. Our research introduced an optimized version of the Swin Transformer by applying three innovative optimization strategies and comparing four distinct model architectures: CNN, Vision Transformer, baseline Swin Transformer, and the proposed optimized Swin Transformer for crop-weed identification performance. The optimized Swin Transformer achieved respectable accuracy alongside the other three models. To evaluate the model's effectiveness across varied farming conditions, the optimized Swin Transformer was retrained on a new and distinct set of crops and weed images. The evaluation showed that the model can adapt well to unfamiliar agricultural contexts and maintain good performance when tested on multiple crop species and farming settings. The primary goal of this project is to minimize the computational cost associated with Swin Transformers without compromising their accuracy and effectiveness through Dynamic Token Clustering (DTC), Adaptive Patch Splitting (APS), and Selective Cross-Scale Attention (SCSA). The optimized architecture demonstrated significant gains across multiple computational metrics while keeping accuracy like that of the baseline model. The lowered processing demands allow for feasible deployment on low-resource agricultural devices like mobile platforms and edge systems.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Fajemisin, Ade UNSPECIFIED
Uncontrolled Keywords:	Precision Agriculture; Crop-Weed Classification; Swin Transformer Optimization; Computational Efficiency; Transfer Learning
Subjects:	S Agriculture > S Agriculture (General) Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	12 Aug 2026 09:07
Last Modified:	12 Aug 2026 09:07
URI:	https://norma.ncirl.ie/id/eprint/9508

Actions (login required)

View Item