Weed Classification for Smart farming in Field Crops: A Machine Learning Evaluation

For any domain, the application of Machine Learning (ML) remains a complicated process that requires experimentation with multiple techniques to find the effective ML model. This process could easily be compromised due to its complexity. In more complicated applications that involve big data and deep learning, it becomes even more important to thoroughly explore the behavior of ML to save time and resources for future applications. This paper explores the application of ML to detect and classify weeds. This is critical as weeds can potentially cause significant financial losses to the agricultural industry. We analyze and compare the performance of 15 ML models for classifying weed types using the widely used ResNet50 and VGG16 pre-trained convolutional neural networks as feature selectors on standard datasets. We conduct two different applications: 1) individual ML algorithms, and 2) stacked generalization which stacks RandomForest, GradientBoosting and SVM, classifying further results with Logistic Regression. For individual ML applications, the summarized results show that RidgeClassifier performed the best on ResNet50 and VGG16, while LinearSVC and LogisticRegression also performed well on both models. However, some models, such as the KNeighborsClassifier and GaussianNB, showed poor performance on both models. The stack algorithm gave the best accuracy of 96.3% with VGG16. Furthermore, the time taken for each model varied greatly, with some models much faster than others. Our detailed results provide guidance regarding the selection of ML models and techniques for weed classification tasks while using pre-trained VGG16 and ResNet50 feature selectors.