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

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.

💡 Research Summary

This paper introduces a novel prototype‑based explainable neural network (ProtoXNet) specifically designed for multi‑channel geospatial data, where each channel corresponds to a distinct physical variable or spectral band. Traditional prototype‑based XAI methods are tailored to RGB images and lack mechanisms to handle the heterogeneous, variable‑specific channels common in climate and remote sensing datasets. ProtoXNet addresses this gap by processing each channel independently through a dedicated encoder that maps the raw channel (size H × W) into a low‑dimensional embedding (size h × w × d). For every channel j, a set of N learnable prototypes Φj (dimension d) is maintained. Similarity between the embedding Z(i)j of a sample i and each prototype is computed using an L2‑based distance, transformed into a normalized similarity map S(i)j (size N × h × w) via a softmax operation. A max‑pooling step extracts the strongest response per prototype, yielding a vector M(i)j ∈ ℝ^N. All channel‑wise vectors are concatenated to form a global feature A ∈ ℝ^{N·C}, which is fed into a linear classifier that outputs class probabilities for K target classes.

Training optimizes a combined loss: cross‑entropy for classification plus a regularization term that encourages prototype diversity and prevents collapse. The architecture remains fully differentiable, enabling end‑to‑end learning.

Two case studies validate the approach. The first classifies phases of the Madden‑Julian Oscillation using six climate variables (e.g., precipitation, 850 hPa temperature, wind components). Ten prototypes per variable capture characteristic spatiotemporal patterns associated with each MJO phase. The second task performs land‑use classification from multispectral Landsat‑8 imagery (eight bands) using fifteen prototypes per band. Prototypes correspond to recognizable spectral signatures such as NDVI peaks for vegetation or high reflectance for water bodies. In both experiments, ProtoXNet achieves classification accuracy comparable to a ResNet‑50 baseline (within 1 % difference) while providing transparent, instance‑level explanations. Visualizing prototype similarity maps reveals that the model’s decisions are driven by physically meaningful patterns, and channel‑wise prototype weights allow quantitative assessment of variable importance (e.g., precipitation dominates certain MJO phases; near‑infrared and green bands are critical for distinguishing vegetation).

The paper also discusses how the prototype framework facilitates global model interpretation, variable relevance analysis, and potential integration into scientific workflows, such as guiding sensor selection or informing climate policy. Limitations include the need to pre‑define the number of prototypes per channel and the computational overhead of maintaining separate encoders. Future work aims to automate prototype count selection, extend the method to temporal sequences, and scale to larger Earth observation systems.

Overall, ProtoXNet bridges the gap between high‑performance deep learning and rigorous, interpretable AI for geospatial learning tasks, offering a scalable solution that respects the multi‑channel nature of environmental data while delivering trustworthy explanations.