1 1 Introduction The yield of cowpea has significantly reduced over the last few years and its decline has been attributed to various factors however the attack by pest and diseases has been severe (Edema,Adipala and Florini, 1997). Disease symptoms in cowpea are so confusing that farmers cannot easily tell if the crop is sick or not. And as such, they end up waiting until in the latest stages when little can be done to save the situation. The current disease control methods in use have not been successful and hence the need for exploring possible alternatives. Accurate diagnosis is the key to mitigating all associated consequences with the transmission of cowpea viruses and bacteria. The most common method of diagnosis adopted by cowpea seed banks is monitoring of viral-like symptoms on plants grown to propagate seed stocks. Other methods that have been used include determining the biological and serological properties of the viruses. At times these methods are used in parallel for greater effectiveness, this intensifies existing efforts and resources devoted to diagnosing disease in cowpea. A study in molecular biology has led to progress of Reverse Transcription Polymerase Chain Reaction (RT-PCR) based methods that facilitate the accurate, rapid and less labor- intensive detection of some of the cowpea infecting viruses (Akinjogunla,Taiwo and Kareem, 2008). All the methods mentioned here require having experts to perform the necessary tasks in diagnosing cowpea diseases. The equipment and chemicals used in the experiments carried out are also very expensive that a smallholder farmer is not able to afford. These methods also take time to deliver results thus the need for a real-time disease diagnosis mechanism that is cheap for the smallholder farmers. New machine learning methods offer an avenue for image recognition and classification models to be easily deployed on mobile devices. Using datasets of plant disease images taken in fields deep convolutional neural network models have been trained to identify plant diseases (Karpathy et al., 2014). In this paper, we discuss automated disease detection model for cowpea based on deep neural network computational techniques that can be used by non-experts and smallholder farmers to do the field-based diagnosis of cowpea diseases. 2 The Cowpea Image Dataset The cowpea leaf images were taken using a form in open data collection toolkit (ODK), the form allowed us to capture the image of the diseased cowpea leaves, assign the disease label based on the symptoms that the leaf showed and it also enabled us to capture the location of the plant from which these leaves were being captured. Data was captured from a total of six districts, four districts from Eastern Uganda and these are Bukedea, Kumi, Ngora, and Serere. More data was collected from Arua district in Northern Uganda and lastly, we collected data from Wakiso district which is in Central Uganda from experimental fields belonging to Kabanyolo Agricultural Research Center under Makerere University in Uganda. A total of 2500 images were collected. The disease classes for which data was collected were Cercospora, powdery mildew, mosaic virus, bacterial blight, scab and healthy. This exercise was done for a period of one-two weeks. All the collected data was uploaded to the server on GoogleApp Engine. Automated image-based diagnosis of cowpea diseases Solomon Nsumba Makerere University Kampala, Uganda snsumba@gmail.com Ernest Mwebaze Makerere University Kampala,Uganda emwebaze@cit.ac.ug Emily Bagarukayo Makerere University Kampala,Uganda ebagarukayo@cit.ac.ug Gilbert Maiga Makerere University Kampala. Uganda gilmaiga@cit.ac.ug Abstract Cowpea is the third most important legume food crop in Uganda with the eastern and northern regions accounting for most of the production in the country. However, it is vulnerable to virus and fungal diseases, which threaten to destabilize food security in sub-Saharan Africa. Unique methods of cowpea disease detection are needed to support improved control which will prevent this crisis. In this paper, we discuss automated disease detection model for cowpea based on deep neural network computational techniques that can be used by non- experts and smallholder farmers to do the field-based diagnosis of cowpea diseases. Image recognition offers both a cost-effective and scalable technology for disease detection. New transfer learning methods offer an avenue for this technology to be easily deployed on mobile devices. Using a dataset of cowpea disease images taken in the field in Uganda, we applied transfer learning to train a deep convolutional neural network to identify three cowpea diseases and to identify healthy plants as well. The best-trained model accuracies were 98% for healthy, 95% for powdery mildew, 98% for cercospora, and 96% for the mosaic virus. The best model achieved an overall accuracy of 93% for data not used in the training process. Our results show that the transfer learning approach for image recognition of field images offers a fast, affordable, and easily deployable strategy for digital plant disease detection. Keywords: Transfer learning, mobile epidemiology, Inception v3 model, MobileNet V1 model