Research and Publications

Introduction

Monitoring inland water quality is essential for ensuring clean water access and protecting aquatic ecosystems. Remote sensing enables large‑scale, frequent observation, but machine‑learning models for water quality prediction face two major challenges: data scarcity and high variability across lakes. Traditional single‑task models struggle to generalize, while existing multi‑task learning approaches either lack flexibility or cannot scale to many diverse lakes with limited samples.

This paper introduces a Multi‑Task Hypernetwork (MTHN) architecture that generates task‑specific model weights from compact embeddings and lake metadata. This approach enables flexible modelling of lake‑level differences while remaining parameter‑efficient and robust to sparse data.

Problems

• Many lakes have very few labeled water‑quality samples

• Lakes differ in depth, temperature, turbidity, and other physical properties

• Current methods trade off parameter efficiency against flexibility

• Large numbers of lakes with few samples each make standard architectures unstable

• Lake‑level metadata is rarely incorporated directly into model weight generation

Method

The Multi‑Task Hypernetwork approach generates lake‑specific prediction models by using a shared hypernetwork conditioned on learned task embeddings and lake metadata. Instead of training a separate model for each lake, the hypernetwork produces the weights of each task‑specific network on demand, allowing the system to capture differences between lakes while remaining highly parameter‑efficient. Metadata such as depth, temperature, or geographic characteristics helps guide the generation of these weights, enabling the model to adapt to diverse environmental conditions even when labeled data is sparse.

Key outcomes

• Improved predictive performance on water‑quality remote sensing tasks.

• Better metadata utilization than existing multi‑task learning baselines.

• Scalability to many lakes with limited samples.

• Superior performance on both water‑quality and other tabular multi‑task datasets.

• Parameter‑efficient architecture suitable for real‑world environmental monitoring systems.

Findings

The study shows that metadata‑driven hypernetworks significantly improve water‑quality prediction across many lakes with limited samples. By combining shared structure with flexible, task‑specific weight generation, the model generalizes better than traditional multi‑task learning methods and handles large numbers of heterogeneous tasks more effectively. The results demonstrate that incorporating metadata directly into the weight‑generation process leads to stronger predictive performance, better scalability, and more robust modeling of complex environmental variation.