Abstract:
Shrimp aquaculture is also faced with some great challenges due to the disease
which can result in 100% mortality rate in shrimp if left untreated. This research
discusses an in-depth assessment of advanced deep learning architectures of
automated classification of diseases in shrimps based on EfficientNet_B3 models
with the addition of calibrated ensemble methods and test-time augmentation (TTA).
We tested eight convolutional neural network architectures that are state-of-the-art
on a medium dataset of images of diseases of shrimp. EfficientNet_B1 and
EfficientNet_B3 were the best models. Subsequently, we came up with advanced
calibrated pipelines consisting of multi stage training, ensemble learning,
temperature scaling and systematic TTA protocols. The accuracy of improved
EfficientNet-B1 was 94% with a good brier score but the Expected Calibration Error
and Maximum Calibration error was poor, while the test accuracy of calibrated
EfficientNet_B3 ensemble was 90.5% with better calibration reliability with a good
brier score, ECE and MCE of 0.0477, 0.1999 and 0.2186 respectively. We use a hybrid
approach of architectural optimization, data augmentation, probabilistic calibration
,ensemble technique and explainable AI to provide robust and reliable disease
classification to practical use in aquaculture. The results reveal that they have made
significant improvements over baseline models and have set new records on
automated detection systems of diseases in shrimps.
