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

TL;DR

The paper tackles the challenge of high-resolution spatial carbon stock estimation from optical remote sensing data. It introduces the Improved Implicit Diffusion Model (IIDM), which fuses knowledge distillation (via KD-VGG and KD-UNet), cross-attention-based feature fusion, and implicit neural representations within a diffusion framework to estimate carbon-stock distribution density from GF-1 WFV imagery. IIDM achieves a leading RMSE of $12.17\%$, representing a $41.69\%$–$42.33\%$ improvement over traditional regression models, and demonstrates substantial reductions in model size and inference time through PCA-based KD and structured pruning. The approach provides high-resolution (16 m) carbon-stock maps suitable for forest-sink regulation and showcases the feasibility of AI-generated content methods for large-scale, data-driven remote sensing applications. The work also validates generalization across regions (e.g., Yunnan Province) and seasons, underscoring IIDM's potential for broad deployment in global carbon monitoring.

Abstract

The forest serves as the most significant terrestrial carbon stock mechanism, effectively reducing atmospheric CO2 concentrations and mitigating climate change. Remote sensing provides high data accuracy and enables large-scale observations. Optical images facilitate long-term monitoring, which is crucial for future carbon stock estimation studies. This study focuses on Huize County, Qujing City, Yunnan Province, China, utilizing GF-1 WFV satellite imagery. The KD-VGG and KD-UNet modules were introduced for initial feature extraction, and the improved implicit diffusion model (IIDM) was proposed. The results showed: (1) The VGG module improved initial feature extraction, improving accuracy, and reducing inference time with optimized model parameters. (2) The Cross-attention + MLPs module enabled effective feature fusion, establishing critical relationships between global and local features, achieving high-accuracy estimation. (3) The IIDM model, a novel contribution, demonstrated the highest estimation accuracy with an RMSE of 12.17%, significantly improving by 41.69% to 42.33% compared to the regression model. In carbon stock estimation, the generative model excelled in extracting deeper features, significantly outperforming other models, demonstrating the feasibility of AI-generated content in quantitative remote sensing. The 16-meter resolution estimates provide a robust basis for tailoring forest carbon sink regulations, enhancing regional carbon stock management.

Figures (8)

• Figure 1: Study area. Notes: (a) China, (b) Yunnan Province, and (c) Huize County. The basemap is provided by ArcGIS Pro.
• Figure 2: Overview.
• Figure 3: PCA-based knowledge distillation framework for VGG. Notes: The PCA-based knowledge distillation process consists of four key steps: (a) global eigenbasis derivation (${{\mathbf{W}}_{\mathbf{N}}},\mathbf{N}=1,2,3,...,16$), (b) blockwise PCA knowledge distillation for feature compression and transfer, (c) the standard autoencoder framework of VGG-19, and (d) the proposed $enc-dec$ autoencoder framework, which enhances efficiency while preserving key structural information. In our method, $relu16$ corresponds to $relu16'$, while the distilled feature representation is denoted as $relu16{_e} = relu16{_d}$.
• Figure 4: Improved implicit diffusion model architecture. Notes: (a) Reverse process of the inference. (b) Denoising model and implicit representation. The denoising model includes KD-VGG feature extractor, encoder of KD-UNet, and conditional network. MLP represents multi-layer perceptron.
• Figure 5: Features of VGG. Notes: Take VGG-19 as an example. Mean explained variance (green area) and mean cumulative explained variance (blue curve) of the $reluN$ features.