The reliability and practical effectiveness of any deep learning-based plant disease detection system arefundamentally determined by the quality, diversity, and representativeness of the dataset on which it is trained.This paper focuses on the systematic construction and preparation of a real-world tomato leaf disease datasetintended to support the development of a Convolutional Neural Network (CNN) based classification model.Unlike studies that employ standardized public repositories, the dataset presented in this work was assembledthrough direct image acquisition from active tomato cultivation fields using mobile camera devices, capturingconditions that accurately reflect the complexity of natural agricultural environments, including variableillumination, background clutter, leaf occlusion, and diverse capture angles.The collected images span multiple disease categories — including early blight, late blight, leaf mold, andbacterial spot — alongside healthy leaf specimens. To ensure the scientific validity of the dataset, all images werereviewed and labeled under the supervision of a certified agricultural officer, whose domain expertise in plantpathology guided accurate and consistent disease classification based on visible symptom indicators. A rigorousquality control process was applied to eliminate ambiguous or low-quality samples.The dataset was further refined through a structured preprocessing pipeline encompassing image resizing to 64 ×64 pixels, grayscale conversion, thresholding-based feature enhancement, and noise reduction. Data augmentationtechniques including rotation, flipping, zooming, and brightness adjustment were applied to enhance datasetdiversity and model generalization. The finalized dataset was partitioned into stratified training, validation, andtesting subsets. This work underscores the critical importance of expert-validated, field-sourced datasets inbuilding robust and deployment-ready agricultural AI systems.
Madhavi Ajay Pradhan (Mon,) studied this question.