MVeLMA: Multimodal Vegetation Loss Modeling Architecture for Predicting Post-fire Vegetation Loss

TL;DR

MVeLMA addresses the challenge of predicting post-fire vegetation loss with uncertainty by integrating nine dynamic meteorological features and twenty-seven static topographical/land-cover features into a multimodal, end-to-end pipeline. It stack-encodes temporal patterns via BiLSTM with attention, models spatial correlations and uncertainty through Gaussian Process Regression, and refines predictions with a Random Forest residual layer, enabling region-specific, confidence-aware vegetation loss forecasts. The approach yields notable improvements over baselines (roughly +10% in R^2 and -11% in NRMSE) and generates county-level confidence maps to inform targeted recovery and policy actions, demonstrated across California counties and cross-region Oregon data. This framework advances practical disaster planning by delivering interpretable, uncertainty-quantified predictions that can guide ecological management and resource allocation under wildfire risk.

Abstract

Understanding post-wildfire vegetation loss is critical for developing effective ecological recovery strategies and is often challenging due to the extended time and effort required to capture the evolving ecosystem features. Recent works in this area have not fully explored all the contributing factors, their modalities, and interactions with each other. Furthermore, most research in this domain is limited by a lack of interpretability in predictive modeling, making it less useful in real-world settings. In this work, we propose a novel end-to-end ML pipeline called MVeLMA (\textbf{M}ultimodal \textbf{Ve}getation \textbf{L}oss \textbf{M}odeling \textbf{A}rchitecture) to predict county-wise vegetation loss from fire events. MVeLMA uses a multimodal feature integration pipeline and a stacked ensemble-based architecture to capture different modalities while also incorporating uncertainty estimation through probabilistic modeling. Through comprehensive experiments, we show that our model outperforms several state-of-the-art (SOTA) and baseline models in predicting post-wildfire vegetation loss. Furthermore, we generate vegetation loss confidence maps to identify high-risk counties, thereby helping targeted recovery efforts. The findings of this work have the potential to inform future disaster relief planning, ecological policy development, and wildlife recovery management.

Figures (4)

• Figure 1: Overview of the proposed pipeline architecture for vegetation loss prediction, combining multimodal data sources, temporal encoding, uncertainty estimation, and predictive modeling.
• Figure 2: Performance analysis across different architectural variations across R2 and MAE metrics. Results highlight the effectiveness of BiLSTM over unidirectional LSTM for capturing temporal fire event and climate dependencies, and RF over MLP for ensemble prediction.
• Figure 3: This figure displays the spatial distribution of MVeLMA's observed, predicted, and predictive confidence estimates across all California counties, where darker colors represent higher vegetation loss and predictive confidences respectively.
• Figure 4: Spatial distribution of MVeLMA's observed, predicted, and predictive-confidence estimates across Butte, California, for four overlapping three-week intervals from July 17 to August 27, 2023. The inset on shows the county’s location within California. The darker colors represent higher vegetation loss and predictive confidence, respectively.