<![CDATA[Manual assessment of microalgae maturity is subjective and does not scale well. This study evaluates whether a convolutional neural network can classify Spirulina microalgae into five visual maturity stages (dead, unripe, under-ripe, medium-ripe, ripe) using color alone. From 5,005 images collected over three months at a single facility, byte-exact duplicates accounted for 41.9% of the dataset; after removal, 2,909 unique images remained. To prevent near-duplicate images from the same session leaking across folds, evaluation used a group-stratified 5-fold cross-validation scheme, with groups defined via perceptual-hash clustering. An ablation across four preprocessing configurations found that combining CLAHE with augmentation, the original proposed method, was numerically weakest (94.74% ± 2.15% accuracy versus 97.43% ± 0.89% with no preprocessing), differing significantly from CLAHE alone (p = 0.021) though not from the plain baseline at this sample size (p = 0.085). A separate comparison, matching fine-tuning depth (last 30 layers) and preprocessing across a custom CNN, MobileNetV2, and ResNet50, found ResNet50 highest at 99.52% ± 0.18% accuracy, significantly ahead of the custom CNN (97.06% ± 1.39%; p = 0.024), while MobileNetV2 showed severe fold-to-fold instability. The custom CNN's advantage lay in efficiency: 54.8 times fewer parameters and roughly 3.5 times faster inference than ResNet50, for only a small accuracy trade-off. A brief, unlabeled field test on six images outside the lab setting showed frequent disagreement among the three models, suggesting limited transfer from controlled to field conditions. Grad-CAM maps are presented as an indication of model attention, not evidence of underlying chlorophyll degradation biochemistry]]>
Copyrights © 2026
