Domain-Robust Marine Plastic Detection Using Vision Models
It addresses the challenge of reliable automation for environmental monitoring by improving cross-domain robustness in underwater debris detection, though it is incremental as it benchmarks existing methods on new data.
This study tackled the problem of domain shift in marine plastic detection by benchmarking vision models on cross-domain underwater imagery, finding that lightweight MobileNetV2 achieved the strongest performance with an F1 score of 0.97, while zero-shot models like CLIP and Gemini showed varying trade-offs in precision and recall.
Marine plastic pollution is a pressing environmental threat, making reliable automation for underwater debris detection essential. However, vision systems trained on one dataset often degrade on new imagery due to domain shift. This study benchmarks models for cross-domain robustness, training convolutional neural networks - CNNs (MobileNetV2, ResNet-18, EfficientNet-B0) and vision transformers (DeiT-Tiny, ViT-B16) on a labeled underwater dataset and then evaluates them on a balanced cross-domain test set built from plastic-positive images drawn from a different source and negatives from the training domain. Two zero-shot models were assessed, CLIP ViT-L14 and Google's Gemini 2.0 Flash, that leverage pretraining to classify images without fine-tuning. Results show the lightweight MobileNetV2 delivers the strongest cross-domain performance (F1 0.97), surpassing larger models. All fine-tuned models achieved high Precision (around 99%), but differ in Recall, indicating varying sensitivity to plastic instances. Zero-shot CLIP is comparatively sensitive (Recall around 80%) yet prone to false positives (Precision around 56%), whereas Gemini exhibits the inverse profile (Precision around 99%, Recall around 81%). Error analysis highlights recurring confusions with coral textures, suspended particulates, and specular glare. Overall, compact CNNs with supervised training can generalize effectively for cross-domain underwater detection, while large pretrained vision-language models provide complementary strengths.