A dual-branch deep learning framework for site-to-pixel upscaling of in situ albedo measurements: Accounting for spatial heterogeneity and scale effects

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Pixel-scale reference ”truth” serves as a critical link between in situ measurements and satellite observations. However, owing to surface heterogeneity, the upscaling of single-site observations to the satellite pixel scale has long been constrained by poorly understood mechanisms of heterogeneity effects, inadequate characterization of nonlinear responses, and insufficient utilization of spatial information. Existing methods, particularly machine learning approaches, generally adopt a direct point-to-pixel mapping strategy that implicitly assumes mutual independence among subpixels within a coarse pixel. This assumption conflicts with the spatial autocorrelation inherent in a spatially continuous land surface, thereby introducing systematic biases into upscaling results over heterogeneous land surfaces. To address this challenge, we propose a site-to-pixel upscaling method that integrates environmental attribute modeling with spatial context learning. The core of this method is a parallel dual-branch architecture comprising a multilayer deep neural network and a convolutional neural network. The former establishes a nonlinear mapping between environmental attributes and the target variable, whereas the latter extracts multiscale spatial heterogeneity patterns within the coarse pixel. The features from the two branches are then adaptively fused through a gating mechanism, followed by spatial aggregation using a point spread function. Evaluation across 416 globally distributed sites encompassing 15 land cover types shows that the model achieves R² values of 0.802–0.997 and RMSE values of 0.007–0.032. At sites jointly affected by spatial heterogeneity and topographic complexity, the median RMSE is more than 30% lower than that of the best-performing benchmark model. Ablation experiments show that the combined mechanism of per-subpixel modeling and residual learning is the primary source of model accuracy improvement, reducing RMSE by more than 60%. The complete dual-branch model accurately reproduces the spatial texture and spatial variability patterns within coarse pixels. Using MCD43A3 product validation as an example, the upscaled reference reduces validation RMSE by 10.2%–22.4%; at highly heterogeneous sites, Bias is reduced by 85.7%. This study demonstrates that the proposed method effectively enhances the ability of in situ observations to represent pixel-scale conditions over heterogeneous surfaces, providing a reliable and cost-effective solution for generating high-quality pixel-scale reference ”truth” and a fundamental basis for remote sensing product validation and the construction of pixel-scale training samples.

This publication has 7 references indexed in Scilit: