Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Minkowski-MambaNet: A Point Cloud Framework with Selective State Space Models for Forest Biomass Quantification

Jinxiang Tu, Dayong Ren, Fei Shi, Zhenhong Jia, Yahong Ren, Jiwei Qin, Fang He

TL;DR

This work tackles the challenge of estimating forest biomass from raw LiDAR point clouds by directly deriving Aboveground Biomass ($\mathrm{AGB}$) and volume without a Digital Terrain Model. It introduces Minkowski-MambaNet, a hybrid architecture that combines sparse 3D convolution (via the Minkowski Engine) with the Mamba Selective State Space Model to capture long-range dependencies, complemented by a Feature Fusion Modification Layer to preserve multi-scale geometry. The Mamba-SEBottleneck enables dynamic, content-aware global context while the skip-connected fusion layer mitigates information loss during downsampling. Evaluated on Danish National Forest Inventory LiDAR data, the method outperforms state-of-the-art approaches in $R^2$ and RMSE for both biomass and volume, demonstrating robustness to boundary artifacts and offering a scalable, DTM-free tool for LiDAR-based forest inventories and monitoring.

Abstract

Accurate forest biomass quantification is vital for carbon cycle monitoring. While airborne LiDAR excels at capturing 3D forest structure, directly estimating woody volume and Aboveground Biomass (AGB) from point clouds is challenging due to difficulties in modeling long-range dependencies needed to distinguish trees.We propose Minkowski-MambaNet, a novel deep learning framework that directly estimates volume and AGB from raw LiDAR. Its key innovation is integrating the Mamba model's Selective State Space Model (SSM) into a Minkowski network, enabling effective encoding of global context and long-range dependencies for improved tree differentiation. Skip connections are incorporated to enhance features and accelerate convergence.Evaluated on Danish National Forest Inventory LiDAR data, Minkowski-MambaNet significantly outperforms state-of-the-art methods, providing more accurate and robust estimates. Crucially, it requires no Digital Terrain Model (DTM) and is robust to boundary artifacts. This work offers a powerful tool for large-scale forest biomass analysis, advancing LiDAR-based forest inventories.

Minkowski-MambaNet: A Point Cloud Framework with Selective State Space Models for Forest Biomass Quantification

TL;DR

This work tackles the challenge of estimating forest biomass from raw LiDAR point clouds by directly deriving Aboveground Biomass () and volume without a Digital Terrain Model. It introduces Minkowski-MambaNet, a hybrid architecture that combines sparse 3D convolution (via the Minkowski Engine) with the Mamba Selective State Space Model to capture long-range dependencies, complemented by a Feature Fusion Modification Layer to preserve multi-scale geometry. The Mamba-SEBottleneck enables dynamic, content-aware global context while the skip-connected fusion layer mitigates information loss during downsampling. Evaluated on Danish National Forest Inventory LiDAR data, the method outperforms state-of-the-art approaches in and RMSE for both biomass and volume, demonstrating robustness to boundary artifacts and offering a scalable, DTM-free tool for LiDAR-based forest inventories and monitoring.

Abstract

Accurate forest biomass quantification is vital for carbon cycle monitoring. While airborne LiDAR excels at capturing 3D forest structure, directly estimating woody volume and Aboveground Biomass (AGB) from point clouds is challenging due to difficulties in modeling long-range dependencies needed to distinguish trees.We propose Minkowski-MambaNet, a novel deep learning framework that directly estimates volume and AGB from raw LiDAR. Its key innovation is integrating the Mamba model's Selective State Space Model (SSM) into a Minkowski network, enabling effective encoding of global context and long-range dependencies for improved tree differentiation. Skip connections are incorporated to enhance features and accelerate convergence.Evaluated on Danish National Forest Inventory LiDAR data, Minkowski-MambaNet significantly outperforms state-of-the-art methods, providing more accurate and robust estimates. Crucially, it requires no Digital Terrain Model (DTM) and is robust to boundary artifacts. This work offers a powerful tool for large-scale forest biomass analysis, advancing LiDAR-based forest inventories.

Paper Structure

This paper contains 17 sections, 11 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional Methods for LiDAR-Based Biomass Estimation
  4. Deep Learning for Point Cloud-Based Biomass Estimation
  5. Method
  6. Sparse Convolutional Network Foundation
  7. Minkowski-MambaNet
  8. Mamba-SEBottleneck
  9. Feature Fusion Modification Layer: Multi-Scale Information Enhancement
  10. Experiment
  11. Data Setup
  12. Evaluation metrics
  13. Experimental Results
  14. Overall Model Performance Comparison
  15. Performance Visualization and Analysis
  16. ...and 2 more sections

Figures (6)

  • Figure 1: In the MSENet50 architecture, there are a total of 16 Mamba-SEBlock, which are distributed across four stages with a layer count of (3, 4, 6, 3). Our strategy involves replacing the last SEBottleneck in each stage with a Mamba-SEBottleneck, resulting in a configuration that includes 4 Mamba-SEBottleneck blocks and 12 SEBottleneck blocks.
  • Figure 2: Histogram of the proportion of coniferous forests in mixed forests across data splits.
  • Figure 3: $R^2$, RMSE, and MAPE of different components of biomass for coniferous and broadleaf trees on the test set. The columns represent$R^2$(the higher the better), RMSE (the lower the better), and MAPE (the lower the better). The corresponding values for wood volume are similar in nature.
  • Figure 4: The test performance plots for the four methods (PointNet, MSENet50, MSENet14, and MinkowskiMamba) each consist of two parts: the biomass residual( left side) and the error distribution ( right side).
  • Figure 5: More examples of AGB subplots with three perspectives in each group: isometric front view, top view, and side view. "High" refers to tall trees, and so on; "Short" refers to short trees; "Single" refers to a single tree; "Small" refers to small trees or sparse forests.
  • ...and 1 more figures