Themeda-Based Spatiotemporal Deep Learning for Predicting Vegetation Dynamics and Fire-Induced Land Cover Change in Northern Australia Using ConvLSTM and Temporal U-Net Models

Abstract

The tropical savannas of Northern Australia, dominated by the perennial C4 grass Themeda triandra (kangaroo grass), are among the most fire-prone and ecologically significant biomes in the Southern Hemisphere. These savannas cover approximately 1.9 million km² and experience annual fire frequencies affecting 20–40% of the total area. Understanding and predicting vegetation dynamics and fire-induced land cover change in these landscapes presents substantial challenges for land managers, conservation practitioners, and carbon accounting frameworks. This study presented a novel spatiotemporal deep learning framework integrating Convolutional Long Short-Term Memory (ConvLSTM) and Temporal U-Net architectures to simulate and map fire-driven land cover change across Northern Australian tropical savannas. The framework incorporated multi-source satellite time series spanning 2019–2022 MODIS NDVI composites (250 m), Landsat 8/9 OLI surface reflectance (30 m), North Australia Fire Information (NAFI) fire scar records, Bureau of Meteorology rainfall grids, and Sentinel-2 MSI data. A novel Themeda Vegetation Index (TVI) was developed from the spectral properties of curing C4 grass tissue and validated against 84 spatially independent field plots (Pearson r = 0.87, p < 0.001 on hold-out subset), outperforming NDVI (R² = 0.621) and EVI (R² = 0.648). Using a spatially blocked train/test partitioning design, 185 non-overlapping 100×100 km geographic tiles were used, ensuring no spatial autocorrelation between partitions. The ensemble model achieved an overall accuracy of 88.9%, Cohen's Kappa of 0.871, and a weighted F1-score of 0.884. NDVI prediction attained a root mean square error (RMSE) of 0.037 NDVI units and a mean absolute error (MAE) of 0.028 NDVI units. The ConvLSTM's forget gate mechanism accurately encoded abrupt fire-induced resets in vegetation states. The Temporal U-Net's learned temporal attention weights identified the dry-season months (June–September) as diagnostically critical for discriminating fire scars without explicit supervision. Post-fire recovery analysis confirmed a strong rainfall gradient: areas receiving>1,200 mm yr⁻¹ recovered to 93% of pre-fire NDVI within 24 months, compared to 57% in transitional low-rainfall zones. Critical limitations were noted. The 2019–2022 training period coincided with a sustained La Niña event (mean Southern Oscillation Index = +11.8 for 2020–2022; Niño 3.4 anomaly = −0.9°C). Recovery rates represented La Niña-period upper estimates, potentially 8–15 percentage points above climatological averages in low-rainfall zones. Model outputs were characterized as enabling spatial technology providing inputs relevant to the Savanna Burning Emissions Reduction (SBER) carbon accounting methodology, subject to project-scale validation by the Clean Energy Regulator. Sub-regional accuracy ranged from 84.1% (Barkly Tablelands) to 91.2% (Top End NT), confirming operationally viable performance across the heterogeneous study area.