Contents lists available at ScienceDirect Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag Original papers Deep learning methods for poultry disease prediction using images George Chidziwisano a ,∗ , Eric Samikwa b , Chisomo Daka c a University of Tennessee, Knoxville, United States b Institute of Computer Science, University of Bern, Switzerland c Malawi University of Science and Technology, Malawi ARTICLE INFO Keywords: Poultry diseases Deep learning Fecal images Disease prediction ABSTRACT As advancements in artificial intelligence continue transforming the world, poultry farming is posed to leverage its benefits with scholars developing deep-learning models for poultry disease prediction. However, existing models have not been evaluated in different contexts to ascertain their generalizability. Further, previous models emphasize multiclass classifiers without considering binary classifiers which could potentially improve prediction accuracy. In this paper, we used fecal data from Tanzania and Malawi to evaluate the generalizability of deep-learning models for predicting poultry diseases. We used data from Tanzania to develop classifiers leveraging MobileNet, DenseNet, and ResNet algorithms. Then, we comparatively evaluated these models using unseen data from Malawi and Tanzania. Our results suggest that MobileNet and DenseNet models performed well on unseen data from Tanzania. Specifically, MobileNet binary, MobileNet multiclass, DenseNet binary, and DenseNet multiclass classifiers achieved an accuracy of 0.98, 0.91, 0.98, and 0.95 respectively. While binary classifiers maintained an accuracy of above 0.8, the performance of multiclass classifiers significantly decreased on unseen data from Malawi. Specifically, MobileNet binary and DenseNet binary classifiers achieved an accuracy of 0.82 and 0.8, while the MobileNet multiclass and DenseNet multiclass classifiers achieved an accuracy of 0.69 and 0.7 respectively. We contribute to previous work by developing the models, evaluating their generalizability, and discussing the challenges of generalizing machine learning models for predicting poultry diseases. 1. Introduction In sub-Saharan African (SSA) countries (e.g., Tanzania, Zambia, Malawi, Kenya, and Nigeria), poultry farming remains a significant source of income and food at the household level (Alders and Pym, 2009). Prior research suggests that the income level of households that practice poultry farming is 2.3 times more than that of households that do not practice poultry farming (Beesabathuni et al., 2018). The difference in income levels between households is attributed to the fact that poultry farmers generate more income when they sell their chickens and eggs. Poultry farming is common among households in rural areas of SSA—in Malawi, 83 percent of the rural households raise poultry compared to 15 percent of the urban households that raise poultry (Assa, 2012). These rural areas are resource-constrained; that is, they tend to have poor road networks, low literate populations, poor communication networks, limited electricity connection, and insuffi- cient social services (veterinary services for their domestic animals). These constraints make it difficult for agricultural and veterinary exten- sion officers to travel to remote areas to monitor the health conditions ∗ Corresponding author. E-mail addresses: gchidziw@utk.edu (G. Chidziwisano), eric.samikwa@unibe.ch (E. Samikwa), cdaka@must.ac.mw (C. Daka). of chickens. Further, these constraints make it difficult to collect large datasets. Artificial Intelligence (AI) has the potential to predict the well- being conditions of chickens in resource-constrained areas. Recently, scholars have been exploring different ways of using Machine Learning (ML) algorithms to predict diseases in chickens (Cuan et al., 2022; Machuve et al., 2022). For example, Machuve et al. used fecal data from Tanzania to develop deep-learning models that can be used to classify chickens that are suffering from Coccidiosis, Salmonella, and New- castle disease (Machuve et al., 2022). They found that deep-learning algorithms, such as transfer learning for image classification, can be used to classify diseases in poultry. Despite this progress, the resulting models were never evaluated against data from a different context to test their generalizability. Further, scholars have prioritized multiclass models without considering binary classifiers as an alternative in cases where multiclass classifiers perform poorly. Given that poultry farming across the globe is not practiced using uniform strategies and different contexts have different conditions that might affect the accuracy of deep learning models, it is important to evaluate existing models with data from other regions. https://doi.org/10.1016/j.compag.2024.109765 Received 3 May 2024; Received in revised form 19 November 2024; Accepted 2 December 2024 Computers and Electronics in Agriculture 230 (2025) 109765 Available online 24 December 2024 0168-1699/© 2024 Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).