IIDM: Improved implicit diffusion model with knowledge distillation to estimate the spatial distribution density of carbon stock in remote sensing imagery

Document Type

Article

Publication Date

2-15-2026

Abstract

Forests are vital carbon sinks, mitigating climate change by reducing atmospheric CO2. Remote sensing enables large-scale carbon stock estimation, but existing methods struggle with computational inefficiency and limited spatial dependency modeling. This study, conducted in Huize County, Yunnan Province, China, using GF-1 WFV (Wide Field View) imagery, proposes the Improved Implicit Diffusion Model (IIDM), a novel deep learning framework integrating knowledge distillation (KD) for enhanced efficiency and accuracy in spatial carbon stock estimation. Our contributions are: (1) A KD-enhanced VGG (KD-VGG) module, reducing inference time by 30 % compared to traditional VGG while preserving critical spectral-spatial features, unlike computationally heavy deep learning models. (2) A KD-UNet with cross-attention and Multi-Layer Perceptron (MLP) modules, improving feature fusion by capturing global-local dependencies, addressing limitations of traditional machine learning models like Random Forest (RF) and Support Vector Regression (SVR). (3) A pioneering IIDM framework combining diffusion modeling with implicit neural representations, offering stable and high-resolution (16-meter) carbon stock estimation, overcoming instability in GAN-based approaches. IIDM achieves an RMSE of 12.17 Mg/ha, a 41.69 %-42.33 % improvement over traditional regression models, enabling efficient, high-precision carbon monitoring for forest management and climate strategies.

Keywords

Carbon stock, Diffusion model, Distribution density, GF-1 WFV, Knowledge distillation

Publication Title

Knowledge Based Systems

ISSN

0950-7051

DOI

10.1016/j.knosys.2025.115131

Funders

Caizhou Technology Remote Sensing AI Innovation Research Fund (CZ-RS-AI-2025-001)

Volume

334

Publisher

Elsevier B.V.

Share

COinS