In marine ecosystems, the presence
of specific underwater plant species such as Gracillaria and Linnaeus is vital
for supporting marine biodiversity and sustaining fishing hotspots. Accurate
detection and classification of these species are crucial for ecological
studies and fisheries management. This research introduces a hybrid object
detection model designed to identify and classify Gracillaria and Linnaeus
species in challenging underwater environments. Given the constraints of
limited and low-quality data, the study utilized two blurry underwater videos,
from which 200 frames were extracted and processed.
The preprocessing stage involved
techniques such as Auto-Orient, Resize (stretching images to 640×640), and
multiple augmentation methods including horizontal and vertical flipping, 90°
rotations, and brightness adjustments. These steps expanded the dataset to
1,043 images, which were then used to train a YOLOv8-based object detection
model. The model achieved a mean average precision (mAP) of 74% and a recall of
66%, indicating its potential in identifying these underwater species despite
the data limitations.
Following object detection, the
identified segments were cropped and further classified using a Convolutional
Neural Network (CNN) model. The CNN was trained on a subset of 412 high-quality
images, which were also subjected to augmentation, resulting in 1,043 images.
The CNN model demonstrated a high accuracy of 98% in classifying the plant
species. When integrated into a pipeline, the combined YOLOv8 and CNN model
system achieved an overall classification accuracy of 87%. Despite these
promising results, the system did exhibit some misclassifications, with five
instances of Linnaeus being detected as Gracillaria, underscoring the need for
further refinement.
The research concludes by
acknowledging the limitations of the current model and dataset, highlighting
the necessity for higher-quality, high-resolution underwater videos to enhance
detection accuracy. Future work will focus on refining the CNN model, possibly
incorporating advanced architectures like DenseNet, and improving the overall
pipeline efficiency to enhance classification performance. This study
contributes to the growing field of marine ecology by providing a foundational
approach for the automated detection and classification of critical underwater
plant species, with potential applications in ecological monitoring and
fisheries management.