查看文章

The problematic and undesirable effects of weeds lead to degradation in the quality and productivity of yields. These unacceptable weeds are close competitors of crops as they constantly devour water, air, nutrients, and sunlight which are helpful for the maturation of crops. For better cultivation and good quality production of crops, weed detection at the appropriate time is an essential stride. In recent years, various state-of-the-art (SOTA) architectures were proposed to detect weeds among crop yields, but they lacked computational cost. This paper mainly focuses on proposing a customized state-of-the-art (SOTA) architecture and comparative study with transfer learning models for detecting and classifying weeds among soybean crops by concentrating on the low computational cost. The selected SoTA is beneficial for detecting weeds on a large scale with very low computational costs. In terms of selection, Maximum Validation Accuracy (MVA), Least Validation Cross-Entropy Loss (LVCEL), and Training Time (TT) were considered for proposing an objective function value system. In total, 15 proposed CNNs with 18 Transfer learning models were analyzed with the help of objective function value and various metric evaluations for finding the best and optimal architecture for weed classification. Experimentation and analysis resulted in C13 being robust and optimal architecture which outperformed every CNNs and Transfer learning model by achieving the highest accuracy of 0.9458 with an objective function value of 5.9335 and ROC-AUC of 0.9927 for the classification of weeds from soybean crops.