Neural Modeling in Agricultural Systems

This paper examines the problem of modeling complex systems using neural modeling. A neural network based on transfer learning of big data was constructed for predicting viral and bacterial diseases in poultry. To optimize the neural network, it was proposed to tune the hyperparameters of the ResNet and VGG 16 models based on modifications to the learning rate and regularization. Disease recognition models for coccidiosis, salmonellosis, and Newcastle disease were constructed with high prediction accuracy, precision, and error. A comparative analysis of models built using transfer learning before and after hyperparameter tuning was conducted. Convolutional neural networks (CNNs) and deep learning models were created for predicting poultry diseases by classifying healthy and unhealthy feces for three disease types. Unhealthy fecal images, which may be symptoms of coccidiosis, salmonellosis, and Newcastle disease, were classified. Models were built using the basic CNN architecture of various configurations. Models were trained using images of bird feces labeled at a farm and during laboratory testing, and then fine-tuned using neural network hyperparameters. The Adam optimizer was used to create the CNN, Resnet, and Efficient Net models for bird disease recognition. Results were obtained for the Resnet model with the Adam optimizer, with and without hyperparameter tuning. Test accuracy results without fine-tuning were 85.06% for the base CNN and 87.85% for the updated neural network with hyperparameter tuning. Fine-tuning with freezing the transfer learning and normalization layers improved the model accuracy to 95.01%, with F1 scores for all classifiers above 83% across all four classes. Prediction reports, confusion matrices, and ROC-AUC curves were plotted for the Efficient Net and Resnet models.