An improved moth flame optimization algorithm based on rough sets for tomato diseases detection

•Feature selection algorithm using the integration of moths-flame optimization and rough set was introduced.•Automatic tomato disease detection approach based on moths-flame optimization and rough set was proposed.•Dataset of tomato leaves infected with disease was collected.•The results demonstrate a good accuracy when MFO and rough set were used together.•The results were evaluated against Accuracy, Precision, Recall and F-Score.

Plant diseases is one of the major bottlenecks in agricultural production that have bad effects on the economic of any country. Automatic detection of such disease could minimize these effects. Features selection is a usual pre-processing step used for automatic disease detection systems. It is an important process for detecting and eliminating noisy, irrelevant, and redundant data. Thus, it could lead to improve the detection performance. In this paper, an improved moth-flame approach to automatically detect tomato diseases was proposed. The moth-flame fitness function depends on the rough sets dependency degree and it takes into a consideration the number of selected features. The proposed algorithm used both of the power of exploration of the moth flame and the high performance of rough sets for the feature selection task to find the set of features maximizing the classification accuracy which was evaluated using the support vector machine (SVM). The performance of the MFORSFS algorithm was evaluated using many benchmark datasets taken from UCI machine learning data repository and then compared with feature selection approaches based on Particle Swarm Optimization (PSO) and Genetic Algorithms (GA) with rough sets. The proposed algorithm was then used in a real-life problem, detecting tomato diseases (Powdery mildew and early blight) where a real dataset of tomato disease were manually built and a tomato disease detection approach was proposed and evaluated using this dataset. The experimental results showed that the proposed algorithm was efficient in terms of Recall, Precision, Accuracy and F-Score, as long as feature size reduction and execution time.