Deep neural networks with transfer learning in millet crop images

Abstract

Plant or crop diseases are important items in the reduction of quality and quantity in agriculture. Therefore, the detection and diagnosis of these diseases are very necessary. The appropriate classification with small datasets in Deep Learning is a major scientific challenge. Furthermore, it is difficult and expensive to generate labeled data manually according to certain selection criteria. The approaches using transfer learning aims to resolve this problem by recognizing and applying knowledge and abilities learned in previous tasks to novel tasks (in new domains). In this paper, we propose an approach using transfer learning with feature extraction to build an identification system of mildew disease in pearl millet. The deep learning facilitates a practically fast and interesting data analysis in precision agriculture. The expected advantage of the proposal is to provide support to stakeholders (researchers and farmers) through the information and knowledge generated by the reasoning process. The experimental result gives an encouraging performance that is the accuracy of 95.00%, the precision of 90.50%, the recall of 94.50% and the f1-score of 91.75%.