OceanCV/European_Green_Crab_Classifier

European Green Crab Classifier

By Ryan Luvera

---

Scientific Context

Correctly documenting the spread of invasive species is incredibly important for stratifying control efforts and predicting future spread.

European green crabs (Carcinus maenas) were first found in Washington state in 1998. Since then, they have invaded into the larger Salish Sea region and have begun to expand into Canada and Alaska.

This invasion is especially worrying due to green crabs impact on native mollusks, crustaceans, and grasses. In Washgington, Dungeness crab fisheries, shellfish aquaculture, and eel grass beds which are critical for juvenile salmon are all in jeporardy due to green crab invasion.

---

Dataset Description

Data were gathered from iNaturalist "Research Grade" category.

Photos were sorted to only include pictures where the carapace was clearly visable (This is a WIP currently, the model includes false data but performs well still).

This meant the removal of pictures that included only the claw, were too blurry to determine distinctive morphology, or were in pieces/broken.

Classes included:

European green crabs and 9 common crab species found in the Salish Sea.

Carcinus maenas
"European green crab"
4044 images
Citation

Cancer productus	Eriocheir sinensis	Hemigrapsus nudus
"Red rock crab"	"Chinese mitten crab"	"Purple shore crab"
884 images	451 images	1537 images
Citation	Citation	Citation

Hemigrapsus oregonensis	Metacarcinus gracilis	Metacarcinus magister
"Yellow shore crab"	"Graceful rock crab"	"Dungeness Crab"
1245 images	177 images	1460 images
Citation	Citation	Citation

Pugettia producta	Romaleon antennarium	Telmessus cheiragonus
"Kelp crab"	"Pacific rock crab"	"Helmet crab"
826 images	1640 images	897 images
Citation	Citation	Citation

Metadata

Each dataset has different sources which can be found within European_Green_Crab_Classifier/metadata. All data were gathered from iNaturalist and the GBIF "Research Grade" photos dataset. Citations can be found within European_Green_Crab_Classifier/citations or by clicking the link below any crab above.

---

Model Selection

YOLOv11 Classification Model

Due to the nature of the data having usually one crab per frame, a YOLOv11n Classification model was trained.

# Train the model

results = model.train(data="/content/output",
                      epochs=400,
                      pateience=50,
                      stream_buffer=True,
                      imgsz=640,
                      hsv_h=0.0,
                      degrees=180,
                      hsv_s=0.5,
                      hsv_v=0.25,
                      translate=0.0,
                      scale=0.25,
                      shear = 2,
                      fliplr=0.5,
                      mosaic=0.0,
                      erasing=0.02,

                      )

Full training arguments can be found within European_Green_Crab_Classifier/model

---

Model Assessment

Confusion Matrix

F1 Score

Precision and Recall

European Green Crab False Negatives

Interpretation

This model exhibits fairly good performance within most classes. The European green crab class resulted in 5% false negatives and Eriocheir sinensis/Hemigrapsus oregonensis being flagged the most for false positives. These results are not super surprising. The poor results from this model are likely due to the data feeding into the model not being refined, and the insufficient training epochs. 20 epochs is much too low, and there are many photos of non-crab species littered throughout the data (as shown above). With the next version of this model, I plan to refine the dataset and create a more refined model that is able to be used with more confidence.

---

Model Use-case

Example Proposal:

Most European green crab invasions are discovered incidentally. Whether it be by citizen scientists, the general public, or trained survey scientists, more standardized methods of survey for green crabs are needed to fully understand dispersal. Remote beaches on invasion fronts of green crab dispersal should be monitored for green crab molt presence. High-costs for survey effort can be mitigated through drone surveys paired with computer vision models. This model is a perfect contender for drones conducting beach crab molt surveys. If green crabs are found, managers can implement more trapping effort to enhance green crab population suppression.

• Study area: Remote beaches in Washington, Canada, and Alaska
• Hypothesis
  • Ho: EGC invasion sites within the fringes of invasion are well accounted for.
  • Ha : EGC molts are more common on remote beaches where trapping isn’t happening heavily.

---

Disclaimer

This model and all data it was trained on are licensed under CC BY-NC 4.0. All data were sourced from iNaturalist and the GBIF. Citations for each of the datasets can be found by clicking on the link below each photo in the above dataset section, or by navitagting to European_Green_Crab_Classifier/citations.