A deep learning framework for mapping evergreen conifer fractional cover at 30 m resolution using fused bi-temporal WorldView and time-series Landsat imagery in mixed mountain forests

Abstract

Evergreen conifers are key components of temperate broadleaf and mixed forests, playing a significant role in shaping ecosystem structure, function, and resilience to climate change. While very high-resolution (VHR) satellite imagery enables accurate classification of evergreen conifers and creation of reference fractional cover maps, scaling this capability to regional levels using coarser-resolution time-series satellite data remains challenging. Traditional machine learning approaches are limited by their inability to fully exploit the spatial detail of VHR imagery and capture sequential patterns in satellite time series. To address these limitations, we developed a deep learning-based framework for mapping evergreen conifer fractional cover at 30 m resolution in mountainous forests. The framework integrates a 3D U-Net model to extract spatial and spectral features from bi-temporal WorldView imagery—while mitigating terrain shadows—and a long short-term memory (LSTM) network to learn sequential dependencies from Landsat time series for regression. We compared our framework against a random forest baseline. Independent spatial and temporal transferability assessments showed that our approach achieved an R² of 0.71 and an RMSE of 0.14, outperforming the benchmark method. To further interpret the spatial predictions, we quantified the spatial configuration of evergreen conifers using landscape metrics across areas with varying conifer cover. Our findings demonstrate the value of combining multi-source, multi-resolution imagery with deep learning models tailored for spatial and temporal complexity. This framework improves the accuracy and transferability of fractional cover mapping and offers a scalable solution for ecosystem monitoring in topographically complex forested landscapes.