Generative Adversial Network (GAN) Synthesized Dataset on Harmful Algal Bloom (HAB)
Description
This attribute set forms a robust framework for detecting Harmful Algal Blooms (HABs), combining key indicators like Bloom_Index, chlorophyll levels, and sea surface temperature anomalies. The parameters cover a range of biogeochemical factors crucial for HAB formation and impact assessment, including nutrient levels (Total_Nitrogen, Total_Phosphorus) and water quality metrics (Dissolved_Oxygen, pH). The system's primary use is in HAB detection and monitoring in coastal waters. However, it has versatile applications: 1. Fisheries Management: Helps predict potential fish kill events by tracking Dissolved_Oxygen levels and Bloom_Index, allowing for timely interventions. 2. Climate Change Research: The Rolling_SST_Anomaly and long-term pH trends can contribute to studies on ocean warming and acidification impacts on marine ecosystems. The model's training using Generative Adversarial Networks (GANs) is a significant innovation. GANs synthesize realistic datasets, addressing the challenge of limited real-world HAB data. This approach: • Expands the training dataset with diverse, simulated HAB scenarios • Improves the model's ability to detect rare or extreme events • Enhances overall prediction accuracy by exposing the model to a wider range of potential HAB conditions By combining real observations with GAN-generated data, the model achieves better generalization and robustness in HAB detection across varied marine environments.
Files
Steps to reproduce
This code implements a GAN-based framework to synthesize realistic Harmful Algal Bloom (HAB) detection datasets. The process begins by defining nine critical marine parameters—including chlorophyll levels, temperature anomalies, and nutrient concentrations—within ecologically valid ranges. A seed dataset (500 samples) is first generated with baseline correlations reflecting real-world relationships, such as the inverse link between chlorophyll and dissolved oxygen, and nutrient-driven chlorophyll increases. The core architecture uses a 4-layer MLP Generator (transforms noise vectors into synthetic samples) and a 4-layer Discriminator (distinguishes real/fake data). After normalizing data to [-1,1], the GAN undergoes 5000 training epochs where both networks iteratively improve—the generator learns to mimic seed data patterns while the discriminator refines its detection ability.
Institutions
- SRM University