Every year, efficient maize production is very important to the economy of many countries. Since nutritional deficiencies in maize plants are directly reflected in their grains productivity, early detection is needed to maximize the chances of proper recovery of these plants. Traditional texture methods recently showed interesting results in the identification of nutritional deficiencies. On the other hand, deep learning techniques are increasingly outperforming hand-crafted features on many tasks. In this paper, we propose a simple transfer learning approach from pre-trained cnn models and compare their results with those from traditional texture methods in the task of nitrogen deficiency identification. We perform experiments in a real-world dataset that contains digitalized images of maize leaves at different growth stages and with different levels of nitrogen fertilization. The results show that deep learning based descriptors achieve better success rates than traditional texture methods. © 2017 IEEE.