01
Independent Research · Environmental Sensing · Machine Learning
HABiSense
A UAV-enabled, multimodal system for earlier harmful-algal-bloom detection.
HABiSense pairs aerial imaging, in-situ water sampling, and a custom deep-learning pipeline to improve early detection of harmful algal blooms. Developed independently, the project encompassed research-question formulation, system architecture, data workflow, model development, evaluation, and technical communication.
Download Paper
System & Significance
Traditional bloom monitoring can separate remote sensing from water-quality measurement, limiting how quickly surface signals are connected to biological risk. HABiSense integrates UAV observation, targeted sampling, multimodal feature processing, and bloom-risk classification in a four-layer workflow.
Core Innovation: The LGFN Prediction Model
LGFN is a parallel dual-branch, multi-task architecture for water-quality time-series data. The local branch applies Conv1D, batch normalization, and ReLU activation to extract short-range anomalies—such as abrupt chlorophyll-a spikes or dissolved-oxygen drops—into a 128-dimensional feature vector. In parallel, a Transformer encoder uses positional encoding and multi-head self-attention to model longer-range dependencies, including nutrient accumulation, seasonal temperature cycles, and delayed phosphorus effects, producing a second 128-dimensional representation.
The two vectors are concatenated into a 256-dimensional representation and compressed through a fully connected fusion layer to 128 dimensions. From this shared feature space, a linear regression head predicts the continuous bloom index, while a linear-sigmoid classification head estimates the probability of bloom presence. The model is trained jointly with mean squared error for regression and binary cross-entropy for classification, allowing both tasks to regularize the shared representation and improve generalization. This local-global fusion helped LGFN achieve an RMSE of 0.0395, an F1 score of 97.85%, an AUC of 0.996, and 32% fewer false negatives than the CNN-LSTM-Attention benchmark.
Five Project Awards
Regional Science & Engineering Fair Category Second Place · Stockholm Junior Water Prize Regional Award · NASA Earth and Space Science Award · United States Air Force Certificate of Achievement · National Ground Intelligence Center Award
