Process-Guided Concept Bottleneck Model

TL;DR

This work tackles the interpretability and trust issues of deep learning in scientific domains by addressing spurious correlations and data sparsity. It introduces the Process-Guided Concept Bottleneck Model (PG-CBM), a generalisation of Concept Bottleneck Models that embeds domain-specific causal processes into the learning architecture, using a multi-branch bottleneck of ecologically meaningful intermediates and a process-grounded aggregator. The authors provide theoretical justification (structural regularisation and causal invariance) and demonstrate empirical gains on Above Ground Biomass Density estimation from Earth Observation data, achieving lower RMSD and bias than vanilla CBMs, a comparable black-box DL model, and existing AGBD products, while delivering interpretable intermediate outputs. The results show improved structure-dependent bias resistance, quantified prediction intervals, and enhanced domain-awareness, arguing that PG-CBM offers a principled path toward trustworthy, science-grounded AI with multi-source supervision. This framework holds promise for broader scientific applications where physical laws and causal mechanisms constrain the mapping from data to outcomes, enabling more robust, transparent, and actionable AI-assisted insights.

Abstract

Concept Bottleneck Models (CBMs) improve the explainability of black-box Deep Learning (DL) by introducing intermediate semantic concepts. However, standard CBMs often overlook domain-specific relationships and causal mechanisms, and their dependence on complete concept labels limits applicability in scientific domains where supervision is sparse but processes are well defined. To address this, we propose the Process-Guided Concept Bottleneck Model (PG-CBM), an extension of CBMs which constrains learning to follow domain-defined causal mechanisms through biophysically meaningful intermediate concepts. Using above ground biomass density estimation from Earth Observation data as a case study, we show that PG-CBM reduces error and bias compared to multiple benchmarks, whilst leveraging multi-source heterogeneous training data and producing interpretable intermediate outputs. Beyond improved accuracy, PG-CBM enhances transparency, enables detection of spurious learning, and provides scientific insights, representing a step toward more trustworthy AI systems in scientific applications.

Figures (7)

• Figure 1: Concept map of the PG-CBM framework. The framework consists of concept and aggregation modules, each implemented as sub-models within the DL architecture. The concept modules predict intermediate biophysical attributes (i.e., ecologically meaningful features that serve as interpretable representations), while the aggregation module combines them to estimate the final target variable.
• Figure 2: Comparison of the PG-CBM (green) with two benchmark DL models: the vanilla CBM (red; see \ref{['sup_section_vanilla CBM']} in the Supplementary Material) and an equivalent black-box model (blue; see \ref{['sup_section_Equivalent black-box DL model']} in the Supplementary Material). Each point represents ground- and model-estimated AGBD values, with a smooth trend line fitted for visualisation. The black dashed line indicates the 1:1 reference. Across the range of predicted AGBD values, PG-CBM exhibits a closer correspondence with the field estimates and reduced spread relative to the benchmarks, which show stronger deviations from the reference line, particularly at higher biomass.
• Figure 3: Comparison of the PG-CBM (green) with ESA CCI biomass map v5.01 (orange) and GEDI L4B product (cyan; see \ref{['sup_section_Other AGBD products']} in the Supplementary Material). Each point represents ground- and model-estimated AGBD values, with a smooth trend line fitted for visualisation. The black dashed line indicates the 1:1 reference. Across the range of predicted AGBD values, PG-CBM exhibits a closer correspondence with the field estimates and reduced spread relative to the CCI and GEDI estimates, which show stronger deviations from the reference line at different biomass levels.
• Figure 4: Comparison of AGBD estimation absolute errors across stem number density (N/ha) for PG-CBM, vanilla CBM, an equivalent black-box DL model, ESA CCI v5.01, and GEDI L4B. Each point represents the AGBD estimation error at a given stem density, with a smooth trend line. Unlike other models showing structure-dependent bias, PG-CBM maintains consistent accuracy across the full density range.
• Figure 5: AGBD predictions from PG-CBM with prediction intervals. Green points show median predictions, grey bars denote the 10–90 percentile interval, the dashed line marks 1:1 agreement, and the green curve summarises the median trend. On average, PG-CBM prediction intervals span approximately $\sim$36±SD% of the median estimate. Note that plot-derived AGBD values themselves also carry substantial uncertainty.