Detection and Classification of Plant Seedlings: Distinguishing Crop Seedlings from Weed Seedlings Using MobileNetV2 Transfer Learning Model with Fine-Tuning Layers

The presented research endeavours to address the notable challenge of distinguishing between crop seedlings and weed seedlings within the disciplines of agriculture and botany. Accurate classification of these categorizations is necessary in order to enhance agricultural practises and optimise crop output. To effectively tackle the issue, the adoption of a comprehensive approach was proposed that involves the use of the MobileNetV2 transfer learning model, together with the integration of fine-tuning layers. The study utilised an extensive dataset of 11,078 annotated images of plant seedlings, encompassing a diverse array of species. In order to enhance the robustness of the model and mitigate imbalances in class distribution, data augmentation techniques were implemented. The dataset was divided into three distinct categories, specifically the training, validation, and testing sets. The training process involved utilising a dataset consisting of 8973 photographs. Subsequently, the model's performance was evaluated by employing a separate set of 1108 images that had not been encountered during the training phase. The model utilised in this study was constructed based on the MobileNetV2 architecture, a well-known framework recognised for its exceptional performance and effectiveness in the field of picture categorization. In order to adapt to the specific attributes of the dataset, custom layers were integrated for the purpose of fine-tuning. In the phase of experimentation, a confusion matrix was utilised to evaluate the performance, resulting in a noteworthy accuracy rate of 97%. This result demonstrates the capacity of the model to effectively decrease the need for manual weed management and the dependence on chemical herbicides in the agricultural sector. Furthermore, our study provides a demonstration of the feasibility of employing deep learning and transfer learning techniques in comparable problems of plant species classification, hence presenting potential avenues for progress in the fields of agriculture and biology. In conclusion, our methodology has effectively showcased a precise approach in differentiating between crop and weed seedlings, therefore facilitating the adoption of sustainable agricultural practises and the protection of the environment.