Multimodal Deep Learning for Hyperspectral Forest Health Monitoring at BIFoR FACE

01Team

MH

Mohammad Hasnain

Student · CMS 462247

MA

Mohid Arshad

Student · CMS 455977

TM

Tahir Mehmood

Student · CMS 458593

Dr. Zuhair Zafar

Advisor

Dr. Muhammad Moazam Fraz

Co-Advisor

02Abstract

Scientists need to know how a mature forest behaves when the air contains more CO₂ than today. The Birmingham Institute of Forest Research (BIFoR) runs a Free Air CO₂ Enrichment (FACE) experiment in which three forest patches receive about 150 ppm extra CO₂ and three matching patches receive ordinary air as controls.

This project proposes a remote sensing and data analysis framework for fixed-camera images covering 25 spectral bands between 665 and 975 nm. It covers data quality checks and calibration, mapping of image regions to experimental patches, calculation of the vegetation indices the sensor can actually support, treatment-versus-control statistics that respect the real experimental design, and simple machine learning models where enough ground truth exists. The final product is a web-based analysis system, CanopyPulse. It does not measure CO₂ directly; it shows how the optical signal of vegetation changes.

03Motivation and problem

Introduction. BIFoR FACE sits in a mature broadleaf woodland in Staffordshire, UK, dominated by roughly 175–180-year-old oak and under CO₂ treatment since April 2017. It has nine patches of about 30 m diameter: three treatment, three ambient-air infrastructure controls and three untouched.

Motivation. Leaf-by-leaf measurement is accurate but slow, while a fixed camera views the same place repeatedly, so it can follow seasonal change and scale up field physiology. Satellites do not fit well, because the patches are only about 30 m wide and the same view of the same arrays is needed over and over.

Problem. The image data cannot be used properly by viewing pictures one by one. The team needs a single repeatable workflow to calibrate raw values, locate the arrays, calculate valid indices, compare treatment and control fairly, and check results against biological data. These steps are not yet joined together. The signal may be small and index-specific — an earlier leaf-level study found that overall reflectance did not change much while one pigment-related index did — so careful, validated analysis is needed rather than a quick classifier.

04Methodology

• Data. About 120 GB of images from two fixed cameras with 25 spectral bands (about 665–975 nm); raw data kept untouched.
• Calibration and QC. Convert raw sensor counts to reflectance using the Empirical Line Method where reference targets exist, and flag data whose calibration cannot be shown to be good.
• Alignment. Register images and map regions of interest to the experimental arrays and dates.
• Features. Spectral statistics and a sensor-compatible subset of a 35-index candidate catalogue (NDVI, NDRE, CIre, NIRv and others).
• Analysis. Treatment versus control compared at array level with mixed-effects or repeated-measures models, effect sizes and confidence intervals, to avoid pseudo-replication.
• ML and validation. Simple, explainable baselines first and only where enough ground truth exists, with leakage-safe validation grouped by image and region, time-blocked and leave-one-array-out.
• Output. The CanopyPulse web system — a Python pipeline, FastAPI backend, React front end and PostgreSQL/PostGIS store.

05Preliminary results

This is a proposal-stage project, so no experimental results are reported yet. Progress so far:

06Expected outcomes

A cleaned, documented inventory of the 2026 dataset; a repeatable calibration and QC workflow, dependent on calibration references; image regions mapped to the experimental units; a validated set of vegetation indices for the 665–975 nm sensor; time plots of spectral change and treatment/control differences where detectable; measured links between selected features and available ground truth; a leakage-aware baseline model where enough labels exist; and a working CanopyPulse prototype with a clear list of limitations.

The main contribution is a repeatable processing pipeline and a usable tool for the BIFoR team.

07Sustainable Development Goals

• 13Climate Action
• 15Life on Land
• 9Industry, Innovation and Infrastructure
• 17Partnerships for the Goals

08Related publications

Key BIFoR publications that this project builds on. These are prior work by the BIFoR FACE community, not outputs of this project.

1. Hart, K. M., et al. (2020). Characteristics of free air carbon dioxide enrichment of a northern temperate mature forest. Global Change Biology, 26(3), 1023–1037. doi.org/10.1111/gcb.14786
2. MacKenzie, A. R., et al. (2021). BIFoR FACE: Water–soil–vegetation–atmosphere data from a temperate deciduous forest catchment, including under elevated CO₂. Hydrological Processes, 35(3), e14096. doi.org/10.1002/hyp.14096
3. Gardner, A., et al. (2021). Is photosynthetic enhancement sustained through three years of elevated CO₂ exposure in 175-year-old Quercus robur? Tree Physiology. doi.org/10.1093/treephys/tpab090
4. Gardner, A., et al. (2022). Are chlorophyll concentrations and nitrogen across the vertical canopy profile affected by elevated CO₂ in mature Quercus trees? Trees. doi.org/10.1007/s00468-022-02328-7
5. Jones, A. L., et al. (2025). The effect of elevated CO₂ on hyperspectral leaf reflectance in mature trees. Trees, 39, 69. doi.org/10.1007/s00468-025-02650-w
6. Kourmouli, A., et al. (2025). Rapid increase in soil respiration and reduction in soil nitrate availability following CO₂ enrichment in a mature oak forest. ACS Omega. doi.org/10.1021/acsomega.4c09495
7. Roberts, A. J., et al. (2022). Effects of elevated atmospheric CO₂ concentration on insect herbivory and nutrient fluxes in a mature temperate forest. Forests, 13(7), 998. doi.org/10.3390/f13070998