BiomSHARP: Biomass Super-resolution for High AccuRacy Prediction

Table of Contents

• Introduction
• Results
  • Performance Comparison of Models Using Sentinel-2 Bands
  • Performance Comparison of Models Using Landsat-5 Bands
• Model Architecture
• Data Description
  • Input Data
  • Ground Truth
  • Data Preparation Steps
• Installation Instructions
• Usage Examples
• Citation

---

Introduction

Accurate estimation of above-ground biomass (AGB) is essential for understanding carbon stocks and flows that inform climate policies. Existing global satellite missions offer valuable environmental monitoring, but their lower spatial resolution limits their application in detailed local assessments.

This repository contains the official implementation of the paper "BiomSHARP: Biomass Super-resolution for High Accuracy Prediction". BiomSHARP is a deep learning model designed to enhance coarse-resolution biomass data from satellite missions, such as ESA's Biomass and NASA's NISAR satellites, by fusing it with high-resolution multispectral data from sensors like Sentinel-2 or Landsat-5. BiomSHARP achieves an output resolution of 25 meters, a fourfold improvement over the original 100-meter resolution.

By bridging the scale gap between global satellite monitoring and local environmental management, BiomSHARP demonstrates superior performance across multiple metrics and outperforms state-of-the-art methods.

📄 You can find the full paper here.

---

Results

Performance Comparison of Models Using Sentinel-2 Bands

Model	opt	bio	Params (↓)	PSNR (↑)	SSIM (↑)	MSE (↓)	RMSE (↓)	MAE (↓)
Bicubic		✔	-	17.25	0.36	7404.83	81.68	56.65
HAT-S		✔	9.6M	21.61	0.49	2531.54	48.58	32.11
ReUse	✔		1.2M	23.07	0.60	1952.91	41.82	27.09
ReUse*	✔		4.9M	23.20	0.61	1895.39	41.19	26.60
SGNet	✔	✔	9.2M	24.42	0.66	1400.91	35.64	22.79
BiomSHARP (ours)	✔	✔	3.4M	24.90	0.70	1254.24	33.70	21.02

Performance Comparison of Models Using Landsat-5 Bands

Model	opt	bio	Params (↓)	PSNR (↑)	SSIM (↑)	MSE (↓)	RMSE (↓)	MAE (↓)
Bicubic		✔	-	17.25	0.36	7404.83	81.68	56.65
HAT-S		✔	9.6M	21.61	0.49	2531.54	48.58	32.11
ReUse	✔		1.2M	23.44	0.62	1820.05	40.24	26.27
ReUse*	✔		4.9M	23.60	0.63	1755.12	39.50	25.70
SGNet	✔	✔	9.2M	24.80	0.69	1280.10	34.09	21.68
BiomSHARP (ours)	✔	✔	3.4M	25.14	0.71	1194.81	32.84	20.45

---

Model Architecture

BiomSHARP combines high-resolution multispectral data with low-resolution biomass data through a guided super-resolution approach. The architecture consists of the following components:

HAT's RHAG module from here:

---

Data Description

Input Data

• High-Resolution Multispectral Data:

  • Sentinel-2: 10-20m resolution bands.
  • Landsat-5: 30m resolution bands.

• Low-Resolution Biomass Data:
  To train BiomSHARP, we use the ESA Biomass Climate Change Initiative (Biomass_cci): Global datasets of forest above-ground biomass. In the future, the intention is to leverage biomass data from the following satellite missions after their planned launches in 2025:

  • ESA's Biomass Satellite
  • NASA's NISAR Satellite

Ground Truth

We utilize high-resolution biomass data as ground truth for model training and validation. These datasets are sourced from the following repositories:

• Eurasia: DOI
• Africa: DOI
• North America (North): DOI
• North America (South): DOI
• South America (North): DOI
• South America (South): DOI
• North Asia (North): DOI
• North Asia (South): DOI
• South Asia: DOI

Data Preparation Steps

The code required for data preparation can be found in the data_preparation folder. It includes:

1. Scripts to download High-Resolution Multispectral Data from Google Earth Engine (GEE).
2. Tools to divide large geographic regions into smaller subimages suitable for model processing.

This preparation ensures the data is compatible with the input requirements of BiomSHARP.

---

Installation Instructions

To set up the environment for BiomSHARP, you'll need the following dependencies:

• CUDA: 11.8
• cuDNN: 8.1
• Python: 3.8.18
• Conda: 22.9.0

Run the following commands to create the environment:

conda create -n biomsharp python=3.8.18
conda activate biomsharp
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install .

---

Usage Examples

Train a model with 4 GPUs:

torchrun --nproc_per_node=4 --master_port=4540 biomsharp/train.py -opt options/train/train_options_biomsharp_landsat.yml --launcher pytorch --auto_resume

Test a model:

python biomsharp/test.py -opt options/test/test_options_biomsharp.yml

---

Citation

@ARTICLE{11265727,
  author={Albors, Laia and Marcello, Javier and Marqués, Ferran},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={BiomSHARP: Biomass Super-Resolution for High Accuracy Prediction}, 
  year={2025},
  volume={63},
  number={},
  pages={1-18},
  keywords={Biomass;Superresolution;Estimation;Data models;Satellites;Biological system modeling;Accuracy;Meters;Deep learning;Forestry;Above-ground biomass (AGB);biomass super-resolution for high accuracy prediction (BiomSHARP);climate monitoring;deep learning (DL);guided biomass super-resolution (SR);multispectral imagery;remote sensing (RS);satellite biomass estimation},
  doi={10.1109/TGRS.2025.3636434}
}