Prototype-based Explainable Neural Networks with Channel-specific Reasoning for Geospatial Learning Tasks

AI summary

Overview

• Research area: Explainable AI (XAI) for geoscience and geospatial machine learning, specifically intrinsically interpretable prototype-based neural networks applied to multi-channel climate and remote sensing data.
• Technical level: Advanced — the paper modifies neural network architecture and training procedures, and assumes familiarity with convolutional encoders, prototype layers, and loss-function design.
• Scope in one sentence: Narayanan and Bergen extend prototype networks so that prototypes are learned separately for each input channel of geoscientific raster data, producing local and global explanations, and demonstrate the approach on a synthetic MNIST-derived task, a Madden-Julian Oscillation (MJO) phase classification task, and a multispectral land-use classification task.

What This Paper Is About

Most explainability tools in the geosciences are post-hoc — they are applied to an already-trained black-box model — and evidence cited by the authors indicates such methods can give inconsistent or conflicting explanations. An alternative is to build interpretable reasoning directly into the network through prototypes: learned patterns, drawn from actual training examples, that a new input is compared against. The problem is that existing prototype networks were designed and benchmarked on RGB natural images, where the three channels jointly encode a single feature (pixel colour). Geoscientific rasters instead have channels that each represent a distinct physical variable or spectral band, so the paper builds a prototype network that reasons about each channel separately.

Key Contributions

1. A channel-specific prototype architecture. The input sample is split channel-wise and each channel passes through a shared encoder to produce its own embedding, which is then compared against its own set of prototypes in a dedicated prototype layer. The model learns N prototypes per class per channel, giving A = N · C total prototypes per class, and the maximum similarity value from each prototype's similarity grid is concatenated across channels into an A-length vector fed to a final linear classifier.
2. A shared encoder across channels, rather than one encoder per channel. Unlike Ghosal and Abbasi-Asl (2021), the authors use a single encoder E with parameters θ_E for all channels, which they state reduces model size and training complexity without loss in model performance (Supplement Section 1).
3. Both local and global explanation mechanisms. Local explanations rank each prototype's contribution score (maximum similarity multiplied by the learned classifier weight linking that prototype to the predicted class) and visualize prototypes via their receptive field using a brute-force "ping" method from Barnes et al. (2022). Global explanations include relative channel importance derived from the final weight matrix ψ_L and the aggregate frequency of high-scoring prototypes across the test set.
4. Three case studies, including multispectral remote sensing. The authors state that applying prototype-based XAI to multispectral remote sensing has not previously been explored, and they also examine whether prototype-based explanations can be derived from pre-trained neural networks whose encoders were originally trained as standard black-box models without prototypes.

Main Findings

• Synthetic MNIST accuracy: the channel-specific prototype network reached 0.978, versus 0.937 for a standard neural network with an identical architecture and 0.685 for a standard prototype network without channel-specific prototypes. Individual class accuracies for the channel-specific model ranged from 0.950 to 1.0.
• Test sample 98 was predicted as class "2", driven by channel 2. The three channels contained the digits "8", "2" and "5"; the highest scoring prototype had a score of 3.890 and matched the "2" digit pattern from training sample 8613 on the right-hand side of channel 2. Prototypes in the other two channels had negligible contribution scores.
• The baseline prototype network without channel-specific prototypes misclassified sample 98, which the authors attribute to its inability to find a jointly learned prototype similar across all channels (Figure S1, Supplement Section 2).
• Global weight inspection matched the synthetic labelling rule. For even-numbered classes, channel 2 prototypes received higher weights and channel 3 prototypes near-zero weights; the pattern reversed for odd-numbered classes; most channel 1 prototypes received near-zero weights.
• Most frequent high-scoring prototypes mostly matched the expected digit patterns, with two exceptions. For even classes, the model identified the matching digit prototype except for class label 8, where the prototype instead captured the "2" digit pattern (possibly due to similarity between "2" and "8"). For odd classes 1, 3, 7 and 9 the matching pattern was found, with the exception of class 5, where a "blank" digit prototype was identified.
• Numeric results for the MJO and land-use case studies are not present in the provided paper content, which is truncated after the synthetic case study's global explanation section. The abstract states only that the method achieves "comparable performance to standard neural networks" across the two geoscientific case studies.
• Baselines used throughout: a standard neural network with an identical encoder architecture and a standard linear classifier, and a standard prototype network without channel-specific prototypes following Barnes et al. (2022), where channels are processed together.

Methodology in Plain English

The architecture has three parts. An encoder turns each input channel into an embedding. A prototype learning layer compares that embedding to a set of learned prototypes — stored as patterns in the embedding space — using a function of the L2 distance (following Chen et al. (2019)), producing a similarity grid for each prototype. An optional location scaling grid ω_j, taken from Barnes et al. (2022), multiplies each similarity map so that similarity is weighted differently depending on where in the grid it occurs. The single maximum value is taken from each prototype's similarity grid, and these maxima are concatenated across all channels and passed to a final linear layer with zero bias and softmax to produce the class prediction.

Training proceeds in three stages. In stage 1 the encoder parameters, prototypes and location scaling parameters are learned while the final layer is frozen; the loss is cross-entropy plus three prototype terms — cluster loss, separation loss and diversity loss (Ming et al., 2019) — computed separately for each channel. In stage 2, a periodic projection step replaces each learned prototype with the closest 1×1×d patch from a real training example, so prototypes remain visualizable; it is performed separately for each channel and at a user-set interval. In stage 3 everything except the final classification layer is frozen and the classifier weights are learned with L1 regularization to encourage a sparser weight matrix. Dropout is used in stage 1 but disabled before projection to keep embeddings stable.

To explain a prediction, the model multiplies each prototype's maximum similarity to the test sample by the classifier weight connecting that prototype to the predicted class, then ranks the results — high-ranking prototypes identify which channel-specific patterns, drawn from which training examples, drove the decision. Global insights come from inspecting the weight matrix for channel-level relevance and from counting which prototypes are most frequently high-scoring across the test set.

Why This Matters

Interpretability is often a precondition for geoscientists to trust a model, and post-hoc attribution methods have been reported to produce inconsistent explanations. By making the reasoning mechanism part of the architecture — and by aligning that mechanism with the fact that geoscientific rasters have physically meaningful, variable-specific channels — this work offers explanations that are traceable to actual training examples rather than to an external attribution heuristic.

Real-world applications:

• Climate monitoring and forecasting: the MJO phase classification case study addresses a phenomenon used in subseasonal-to-seasonal outlooks.
• Land-use and land-cover mapping from multispectral satellite imagery, the third case study.
• Environmental and Earth observation pipelines more broadly, where users need to know whether a model is relying on physically sensible variables.
• Scientific model auditing, since global explanations can reveal whether a channel is being ignored (near-zero classifier weights) or over-used.

Industry relevance: operators of Earth observation and geospatial analytics platforms, climate risk and agricultural monitoring services, and any organisation deploying deep learning on multi-band raster data where regulatory or scientific review requires justification of automated decisions.

Future Directions

• Report and analyse the MJO and land-use case study results in full, including whether channel-specific prototypes change which variables the model relies on relative to a standard network, and how the explanations compare to post-hoc methods such as SHAP and LRP.
• Determine how well prototype-based explanations can be recovered from encoders pre-trained as standard black-box models, an extension the authors state they examine in this work but whose outcomes are not visible in the provided content.
• Investigate the observed failure modes, such as the class-8 prototype capturing the "2" digit pattern and the class-5 prototype capturing a "blank" pattern, to understand whether these stem from genuine pattern similarity or from limitations of the similarity and diversity losses.
• Address the two geoscience-specific gaps the authors flag as unaddressed by existing prototype methods: highly correlated features and multi-scale spatiotemporal patterns, beyond the location scaling already borrowed from Barnes et al. (2022).

Target Audience

Researchers and practitioners working at the intersection of machine learning and the Earth sciences — particularly those applying deep learning to multi-channel raster or gridded data who need interpretable models rather than post-hoc explanations. It is also relevant to explainable-AI researchers interested in domain-specific adaptations of prototype networks, and to applied scientists and engineers in climate, remote sensing and geospatial analytics who must justify model predictions to domain experts or reviewers. The paper is a methods contribution published in Environmental Data Science; readers without a background in neural network architecture will find the technical sections demanding.

Authors’ abstract

Explainable AI (XAI) is essential for understanding machine learning (ML) decision-making and ensuring model trustworthiness in scientific applications. Prototype-based XAI methods offer an intrinsically interpretable alternative to post-hoc approaches which often yield inconsistent explanations. Prototype-based XAI methods make predictions based on the similarity between inputs and learned prototypes that represent typical characteristics of target classes. However, existing prototype-based models are primarily designed for standard RGB image data and are not optimized for the distinct, variable-specific channels commonly found in geoscientific image and raster datasets. In this study, we develop a prototype-based XAI approach tailored for multi-channel geospatial data, where each channel represents a distinct physical environmental variable or spectral channel. Our approach enables the model to identify separate, channel-specific prototypical characteristics sourced from multiple distinct training examples that inform how these features individually and in combination influence model prediction while achieving comparable performance to standard neural networks. We demonstrate this method through two geoscientific case studies: (1) classification of Madden Julian Oscillation phases using multi-variable climate data and (2) land-use classification from multispectral satellite imagery. This approach produces both local (instance-level) and global (model-level) explanations for providing insights into feature-relevance across channels. By explicitly incorporating channel-prototypes into the prediction process, we discuss how this approach enhances the transparency and trustworthiness of ML models for geoscientific learning tasks.