DeepWatch

Inspiration

Sea turtles and dolphins are really cute. Thousands of them are also needlessly killed every day.

Commercial fishermen often end up catching marine animals they didn't intend to catch, such as sea turtles, dolphins, seals, sharks, and juvenile fish, discarding them after catching them. This is bycatch. A prominent issue, bycatch, threatens endangered species and destroys marine ecosystems. Currently, there are organizations that are helping fisheries manage unwanted catch, but our machine learning approach assesses risk levels of bycatch for anyone with just a click of a button.

What it does

We created an XGBoost model to predict the likelihood of bycatch being present at a given location, such as sea turtles, dolphins, seals, sharks, and juvenile fish. Based on the latitude, longitude, sea surface temperature, current speed, current direction, hour of day, migration pattern, target species, and the species' fate (whether it was kept on the ship or discarded back into the ocean), the model decides whether bycatch is a risk to fishermen or not.

How we built it

Machine Learning Model

Data

Simulated Data - Due to limitations of features within open-source datasets, a dataset was simulated with slight noise to train a model for future use in actual datasets.

Features Used (9 total)

Latitude
Longitude
Sea Surface Temperature
Current Speed
Current Direction (encoded)
Hour of Day
Migration Pattern (encoded)
Target Species (encoded)
Species Fate (encoded)

Algorithm

Gradient Boosting Classifier - An ensemble method that builds multiple decision trees sequentially, with each tree correcting errors from previous ones. An efficient and fast algorithm that provides high accuracy, as opposed to some more complex algorithms that have a boot time greater than what is allowed for the Render free plan.

Model Configuration

n_estimators=200      # 200 decision trees
learning_rate=0.1     # Step size for corrections
max_depth=4           # Maximum tree depth
min_samples_split=10  # Minimum samples to split a node
subsample=0.8         # 80% data sampling per tree
random_state=42       # Reproducible results

Training Process

Data Loading: Reads Excel file
Label Creation: Converts "Present"/"Absent" to 1/0
Encoding: Transforms categorical variables (direction, species, etc.) to numeric
Train/Test Split: 80% training, 20% testing (stratified)
Model Training: Fits Gradient Boosting model
Persistence: Stores model and encoders in memory

Risk Classification

High Risk: Probability ≥ 65%
Low Risk: Probability < 65%

Tech Stack

All hosted on Render (free version leads to slight loading issue with long periods of inactivity)

Backend

FastAPI Backend

Getting values from input fields in the frontend
Applying the machine learning algorithm to get the probability
Providing this output to the frontend

Frontend

Simple HTML/CSS/JS frontend

Landing page with information about the DepepWatch initiative
Risk Predictor page with all input fields and submission field that outputs percentage on bars and metrics from the model

Challenges we ran into

Finding a real-world dataset with the combination of features we needed was genuinely difficult. Most publicly available bycatch datasets are either poorly formatted, heavily aggregated, or restricted to categorical variables that cannot support numerical model training. We simulated our training data from known IOTC catch patterns, which gave us clean numerical inputs but introduced overfitting risk, since validation was performed on data drawn from the same distribution. We were deliberate about acknowledging this limitation and structured the model to be straightforwardly retrained once real observational data becomes available. Also, as we used the free version of Render to host our webpage, after a period of inactivity, it restarts the building process, which makes the webpage take some time to load before redirecting the user to it.

Accomplishments we are proud of

Within a single day, we built a functioning end-to-end system: a trained machine learning model, a REST API backend, and an interactive frontend, all connected and deployed to the web. The fact that a fisherman can open a browser, enter their fishing conditions, and receive a data-driven risk assessment in under a second felt like a genuine proof of concept for what accessible, practical environmental intelligence could look like in the real world.

What we learned

We learned how to architect a full-stack machine learning application from scratch, covering data simulation, model training, API design with FastAPI, frontend integration, and cloud deployment on Render. We also developed a clearer understanding of the real-world constraints around environmental datasets and why domain-specific data collection is often the hardest part of any conservation-focused machine learning project. Working under time pressure using GitHub reinforced the value of clean separation between system components so teammates could build in parallel without blocking each other.

What's next for DeepWatch

We plan to grow DeepWatch in three directions. First, we will expand species coverage beyond the current five to include the full range of commercially targeted species in North Carolina and broader Atlantic waters. Second, we will integrate real-time ocean data feeds for sea surface temperature, current conditions, and migration tracking so that fishermen no longer need to manually input environmental parameters. Third, we will build spatial visualizations showing where at-risk marine animal populations are currently located relative to a vessel's planned fishing route, transforming DeepWatch from a risk calculator into a genuine navigation tool for sustainable fishing.