Models for recognition and classification of agricultural weeds based on transfer learning

Neural network models for recognizing and classifying weeds based on deep learning have been built. Classification models for the main, most common weeds have been created. The created classification models are based on the use of transfer deep learning VGG19, DenseNet, MobileNetV2 and ResNet152. Modern Adam deep learning optimizers for a given model architecture were used to optimize neural networks. An open data system was used to train these models, and our own data set collected from agricultural land was used for testing. Three classes of weeds are classified. Which are the most common in agriculture. When testing the model, the belonging of the plant to one of the classes of sorghum, cereal weeds and broadleaf weeds is determined. The weed itself and the prevalence of this weed plant are determined. A matrix of forecasting errors is constructed in quantitative and percentage terms of correctly and incorrectly classified classes. A full report of forecasting the classification of weed plants by classes for the completeness of the model in percentage terms, which indicate the ability of the model to classify and identify all data in the corresponding class, is created. The accuracy of the model is determined, which is capable of classifying and returning only data belonging to a given class. Also, estimates of a single metric F-1 are built, which combines completeness and accuracy using the harmonic mean. A full comparative analysis of all four classes of models for the classification of weed plants is carried out and the most accurate for each of the trained models are indicated in percentage and numerical form. Roc curves and predicted values of the classification of weed plants are built. The built models are based on a modern method for constructing complex models - transfer learning.