Weed Crop Identification and Classification Using Transfer Learning with Variants of MobileNet and DenseNet

Weed, often referred to as unwanted plants that grow in agricultural fields, poses a major challenge in modern farming, resulting in reduced productivity. Our project aims to tackle this issue by exploiting deep learning techniques. Weeds like Kochia, Ragweed, Horseweed, and Redroot Pigweed (RRPW) continue to be a major obstacle to agricultural productivity, competing with crops for essential resources. This study explores the effectiveness of deep learning models in identifying and classifying these weed species. Specifically, it compares the performance of MobileNetV1, MobileNetV2, and three DenseNet variations (DenseNet-121, DenseNet-169, DenseNet-201).A dataset comprising 5,148 weed images was utilized, with data augmentation techniques applied to enhance model training. Among the models tested, DenseNet-169 achieved the highest accuracy at 97%, demonstrating exceptional capability in distinguishing weed species. DenseNet-121 and DenseNet-201 also performed well, with testing accuracies of 95% and 96%, respectively. MobileNetV2 achieved a testing accuracy of 92%, emphasizing its computational efficiency while maintaining strong performance across most weed species. To provide a comprehensive evaluation, both accuracy and F1-score are reported. Accuracy gives an overall measure of correctness, while F1-score is crucial for class imbalance, ensuring reliable classification across different weed species. Notably, DenseNet-169 achieved F1 scores of 0.98, 0.96, 0.97, and 0.97 for Kochia, Ragweed, Horseweed, and Redroot Pigweed, respectively, further highlighting its effectiveness.