GREAT: Generalizable Representation Enhancement via Auxiliary Transformations for Zero-Shot Environmental Prediction

TL;DR

GREAT addresses the challenge of predicting environmental dynamics in unmonitored regions by learning generalizable representations through auxiliary, multi-layer data transformations. The method uses a model-agnostic framework with an LSTM predictor and learnable transformations applied to inputs and temporal representations, optimized via a bi-level training regime guided by sparsely labeled reference domains and reconstruction constraints. Key contributions include the formulation of $g_ ext{input}$ and optional $g_ ext{hidden}$ transformations, a bi-level objective that leverages $\,\mathcal{L}_\text{lower}$ and $\,\mathcal{L}_\text{upper}$ with reconstruction losses, and a comprehensive pre-training strategy, all validated on six watersheds for zero-shot stream temperature prediction. The results show that GREAT consistently outperforms baselines, improves temporal fidelity, and generalizes augmented data to diverse architectures, offering a practical tool for environmental monitoring where dense monitoring is infeasible.

Abstract

Environmental modeling faces critical challenges in predicting ecosystem dynamics across unmonitored regions due to limited and geographically imbalanced observation data. This challenge is compounded by spatial heterogeneity, causing models to learn spurious patterns that fit only local data. Unlike conventional domain generalization, environmental modeling must preserve invariant physical relationships and temporal coherence during augmentation. In this paper, we introduce Generalizable Representation Enhancement via Auxiliary Transformations (GREAT), a framework that effectively augments available datasets to improve predictions in completely unseen regions. GREAT guides the augmentation process to ensure that the original governing processes can be recovered from the augmented data, and the inclusion of the augmented data leads to improved model generalization. Specifically, GREAT learns transformation functions at multiple layers of neural networks to augment both raw environmental features and temporal influence. They are refined through a novel bi-level training process that constrains augmented data to preserve key patterns of the original source data. We demonstrate GREAT's effectiveness on stream temperature prediction across six ecologically diverse watersheds in the eastern U.S., each containing multiple stream segments. Experimental results show that GREAT significantly outperforms existing methods in zero-shot scenarios. This work provides a practical solution for environmental applications where comprehensive monitoring is infeasible.

Figures (7)

• Figure 1: The overall framework of GREAT.
• Figure 2: Map of six distinct and ecologically varied watersheds in the eastern United States.
• Figure 3: Difference of RMSE between GREAT and DANN. Red color means GREAT is better.
• Figure 4: Comparison of feature importance attributions for DANN and GREAT.
• Figure 5: Time-series comparison of stream water temperature predictions by GREAT, DANN and LSTM.