Wildfires constitute one of the most significant drivers of biodiversity loss and ecosystem degradation at the global scale. Accurate prediction of post-fire vegetation recovery rate is critically important for ecological restoration planning. Conventional single-task (STL) machine learning models treat the vegetation recovery index as an independent target variable, thereby neglecting the reciprocal interactions between recovery dynamics and environmental covariates such as soil moisture and slope. This study systematically evaluates the performance advantage of multi-task learning (MTL) over single-task models in predicting post-fire vegetation recovery rate. The Normalized Difference Vegetation Index (NDVI) recovery rate is defined as the primary task, while soil moisture and terrain slope are simultaneously modeled as auxiliary tasks. Hard-parameter sharing and soft-parameter sharing MTL architectures were developed using multi-temporal remote sensing data from Sentinel-2 and ERA5-Land datasets. Experimental results demonstrate that the hard-parameter sharing MTL model achieves a 14.3% improvement in mean squared error (MSE) and a 0.08-point increase in the coefficient of determination (R²) compared to its single-task counterpart. Furthermore, gradient conflict analysis reveals that the soil moisture task establishes a positive transfer relationship with the primary task. These findings suggest that multi-task learning frameworks represent a promising tool for post-fire ecosystem monitoring and early warning systems.
Kaan Alper (Wed,) studied this question.