A fast and accurate expert system for weed identification in potato crops using metaheuristic algorithms

Precision Agriculture is an area that can benefit from the latest advances in expert systems. One of the goals is to detect and remove weeds intelligently, so that herbicides are only sprayed in areas where weeds exist. By reducing the amount of herbicides used, the risk of contamination of crop and water resources is avoided, which could produce harmful environmental effects. In this paper, a new computer vision based expert system is presented for identifying potato plants and three different kinds of weeds (Secale cereale L., Polygonum aviculare L. and Xanthium strumarium L.) in order to perform site-specific spraying. The videos were recorded from two Marfona potato crops, with a total area of 4 ha located in Kermanshah-Iran (34°20′17.203′′N, 46°19′56.807′′E), taken with a moving platform with a speed of 0.13 m/s, under outdoor lighting conditions. Applying image processing, 3459 objects were extracted and used to train and test the classifiers. 126 color features and 60 texture features were extracted from each object. The main contribution of the proposed approach was the application of two metaheuristic algorithms to optimize the performance of a neural network classifier: first, the cultural algorithm is used to select the five most effective features, in order to improve computational efficiency; then, the harmony search algorithm is applied to find the optimal configuration of the network. This approach has been compared with a statistical method based on linear discriminant analysis. The experimental results show that the proposed expert system achieves an excellent identification accuracy of 98.38%, requiring less than 0.8 s of execution on an average PC.