×
The submission system is temporarily under maintenance. Please send your manuscripts to
Go to Editorial ManagerIdentifying fish species in natural aquatic environments remains challenging due to changing light conditions, turbid water, and complex underwater scenes. Most current deep-learning models rely on controlled datasets, which limits their use in real-world settings. This study presents Auto Fish, a mobile deep-learning system for real-time, offline fish species identification on Android devices. The system uses the MobileNetV2 architecture, optimized with TensorFlow Lite for processing on the device. This approach ensures high accuracy while keeping computational costs low. We trained and evaluated the model on a balanced dataset of 8,000 annotated images, including nine marine species: Sea bass, Red sea bream, Horse mackerel, Gilt-head bream, Shrimp, Black sea sprat, Trout, Red mullet, and Striped red mullet. Extensive preprocessing, image enhancement, and stratified sampling helped the model perform well despite variations in lighting and background conditions. The experimental results showed a validation accuracy of 99.2%, with both macro and micro Precision, Recall, and F1-scores around 99.3%, and an average False Positive Rate (FPR) of 0.09%. The system supports offline recognition, cloud syncing via Firebase, and delivers real-time results within 4.2 seconds per image on mid-range smartphones. These findings show that Auto Fish can effectively classify fish species in the field while remaining efficient and easy to use. This work offers a practical AI-based solution that connects research with ecological monitoring, empowering citizen scientists and conservationists to document biodiversity using mobile technology.
In precision agriculture, crop disease detection can be a highly valuable undertaking in which scalable and correct solutions may save considerable amounts of money and loss of yield. This paper introduces a comparative analysis of state-of-the-art deep learning models with special attention to EfficientNetB3 hybrids, which are trained on a balanced subsample of the PlantVillage dataset with 33 classes based on nine crops. To overcome the shortcomings of the previous studies, which used unbalanced sample, a leakage-free balancing approach was used, resulting in 13,200 training and 3,300 validation samples. Custom head transfer learning was used where it was tested using two strategies; FreezeUnfreeze fine-tuning, and Singlephase training. MobileNetV2, InceptionV3, DenseNet121, GhostNet, in addition to other baseline CNNs, were compared to baseline Convolutional Neural Networks (CNNs). The findings indicate that EfficientNetB3 hybrids are superior with an accuracy of ≥99.5% and 99.9% Area Under the Curve (AUC) and specificity than the previous CNN-based systems. The paper logically defines a performance ladder between model options and real-life deployment demands, such as lightweight mobile applications to precision agriculture systems, and points out future trends in the field-based validation.