Preprint
Article

This version is not peer-reviewed.

Spaceborne GEDI LiDAR and Deep Learning Quantify Edge Driven Aboveground Biomass Losses in Madhupur Sal Forests

Submitted:

11 November 2026

Posted:

13 November 2026

You are already at the latest version

Abstract
Tropical forest fragmentation is a critical driver of carbon loss, yet edge‑driven biomass degradation remains poorly quantified across South Asia. This study introduces a GeoAI fusion framework that integrates spaceborne LiDAR from NASA’s GEDI Level 4A footprints with Sentinel‑2 multispectral imagery to model above‑ground biomass density (AGBD) across the 18,439.58 hectares Madhupur Sal (Shorea robusta) forest landscape in Bangladesh. GEDI tracks were filtered using strict waveform quality criteria (l4_quality_flag = 1; algorithm_run_flag = 1) and fused with spectral indices and a continuous Euclidean distance‑to‑edge matrix. A regularized PyTorch Multi‑Layer Perceptron with LeakyReLU activations was trained to predict biomass at 10‑meter resolution, achieving strong performance (R² = 0.78; RMSE = 21.6 Mg/ha). By integrating vertical structural metrics from GEDI, the model overcame passive optical saturation limits up to 203.6 Mg/ha. Wall‑to‑wall projections revealed pronounced spatial heterogeneity: intact forest cores stored maximum carbon, while edge zones declined sharply. Edge‑effect profiling indicated biomass losses penetrating ~90 meters into interiors, rendering patches narrower than 180 meters structurally compromised. These findings provide the first quantitative evidence of edge‑driven carbon depletion in Bangladesh’s Sal forests. Beyond ecological insight, the study demonstrates a cloud‑scalable, reproducible GeoAI workflow for monitoring biomass dynamics across fragmented tropical landscapes. Incorporating fragmentation geometry and edge‑effect parameters into national greenhouse gas inventories and REDD+ verification frameworks is urgently needed. The proposed approach highlights the potential of combining LiDAR and machine learning to advance forest monitoring and conservation in data‑limited regions.
Keywords: 
;  ;  ;  ;  ;  ;  
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.