Optimized convolutional neural networks for real-time detection and severity assessment of early blight in tomato (Solanum lycopersicum L.)

Abstract

Early blight, caused by Alternaria alternata, poses a critical challenge to tomato (Solanum lycopersicum L.) production, causing significant yield losses worldwide. Despite advancements in plant disease detection, existing methods often lack the robustness, speed, and accuracy needed for real-time, field-level applications, particularly under variable environmental conditions. This study addresses these gaps by leveraging transfer learning with optimized MobileNet architectures to develop a highly efficient and generalizable detection system. A diverse dataset of 6451 tomato leaf images, encompassing healthy and varying disease severity levels (low, medium, high) under multiple lighting conditions, was curated to improve model performance across real-world scenarios. Four MobileNet variants—MobileNet, MobileNet V2, MobileNet V3 Small, and MobileNet V3 Large—were fine-tuned, with MobileNet V3 Large achieving the highest classification accuracy of 99.88 %, an F1 score of 0.996, and a rapid inference time of 67 milliseconds. These attributes make it ideal for real-time IoT applications, including smartphone-based disease monitoring, automated precision spraying, and smart agricultural systems. To further validate diseased samples, internal transcribed spacer (ITS) sequence analysis confirmed A. alternata with over 98 % similarity to known isolates in the NCBI database. This study bridges critical research gaps by providing a robust, non-destructive, and real-time solution for early blight severity assessment, enabling timely, targeted interventions to mitigate crop losses in precision agriculture.