Classification of common diseases in agaricus bisporus by deep learning method

In this thesis study, a deep learning-based system is proposed for the image-based classification of four common diseases observed in Agaricus bisporus (white button mushroom), along with the healthy condition. The five classes considered are: healthy, bacterial blotch, cobweb (white mold), dry bubble, and wet bubble. Using a custom-designed imaging cabin, image datasets were created under controlled conditions for three distinct lighting setups: white light, ultraviolet (UV) light, and mixed (white + UV) light. A total of 20 different pretrained convolutional neural network (CNN) architectures were systematically evaluated. Each model was assessed using various performance metrics, including accuracy, recall, precision, F1-score, ROC AUC, PR AUC, specificity, Matthews correlation coefficient (MCC), log loss, and a composite performance index (CPI). The CPI was calculated as a weighted average of five key metrics to support holistic decision-making. Results indicate that ResNet-50 consistently achieved the highest performance across all lighting conditions. While white light yielded the best overall class separation, mixed light provided balanced metric scores in some cases, and UV light introduced increased classification uncertainty in certain classes. Class-based evaluations revealed that the models achieved high accuracy in identifying the healthy and cobweb classes, whereas bacterial blotch and wet bubble classes posed challenges due to visual similarities. Confusion matrices, along with ROC and PR curves, enabled detailed visual analyses, highlighting the distinct strengths and weaknesses of each model across different classes. These findings underscore the significant impact of both CNN architecture depth and the lighting conditions used during data collection on classification performance. In conclusion, this study confirms the potential of image-based disease detection in mushroom cultivation and highlights the suitability of the proposed system for early diagnosis, automated monitoring, and smart agriculture applications.