Published Paper Details:

Identifying weeds in a timely manner is essential to ensure there is no reduction in the overall yield in contemporary agriculture. Timely identification of weeds leads to increased efficiency, which can minimize weeds and herbicide use, and increase productivity. In chili cultivation, weeds have the potential to look similar to the crops, which makes it labor-intensive, error-prone, and inefficient to do weed detection manually. For this study, the data were drawn from three models with deep-learning algorithms to demonstrate a method for weed detection in chili. The three models that were utilized were: ConvNeXt, Vision Transformer (ViT), and Swin Transformer. We created a custom image dataset from the real-world chili farmlands, where we collected images of the crops with many difficulties including multiple lighting conditions, occlusion, and other background trouble. It was necessary to conduct a lot of pre-processing and augmentations to ensure the models had the best possibility to be trained and generalized properly. The models were fine tuned using transfer learning on the models based on a binary classification model (weed vs non-weed). The possible effects of models were evaluated based on their accuracy, precision, recall, f1 score, and inference time. ConvNeXt achieved the highest accuracy at 99.0%, followed by ViT at 97.8%, and Swin Transformer at 92.1%. ConvNeXt also had the highest inference time and generalization properties, making it applicable on a small scale for real-time applications on low-power devices. Overall, the study's findings demonstrate that deep learning solutions based on modern transformer models and hybrid CNN architectures could fundamentally alter how systems for automated weed control have been introduced and how many herbicides are needed to conduct a productive grow as we find ourselves working towards reducing our dependency for chemical solutions.