The study, published in Climate Dynamics by Ziyu Jiang of Beijing Normal University and the Potsdam Institute for Climate Impact Research, together with Ming Wang, Kaiwen Li, veteran complexity scientist Jürgen Kurths, and Kai Liu, builds on a decade of work applying network science to the Earth system. In the climate network framework, each grid point of a atmospheric field becomes a node, and statistical relationships between the time series at different locations become edges connecting those nodes. When two regions of the atmosphere fluctuate in a coordinated fashion—whether through shared local weather or genuine long-range teleconnections—an edge forms between them. The resulting graph can then be interrogated with the standard toolkit of network theory: degree, clustering, path lengths, and other topological metrics that quantify how the system is organized.

What makes the new work distinctive is its focus on mean sea level pressure, the very field that tropical cyclones disturb most directly. A developing cyclone is, at its core, a dramatic reorganization of the pressure field: air spirals inward toward a deepening low-pressure center, and the surrounding pressure gradients steepen and shift on timescales of hours to days. The researchers hypothesized that this reorganization should leave a detectable imprint on the structure of a network built from pressure anomalies. To test the idea, they constructed evolving undirected networks from six-hourly mean sea level pressure anomalies over the Western North Pacific, using rolling ten-day windows so that the network continuously adapts to the changing state of the atmosphere. Edges between nodes were established based on two criteria: temporal coherence, meaning the pressure series at two locations fluctuate in step, and magnitude similarity, meaning the strength of their variations is comparable.

The results are striking. When a tropical cyclone is present in a given ten-day window, the network metrics do not change randomly—they change in organized, spatially coherent patterns. The team examined four network measures and found that they trace out low-value bands aligned with the storm’s track, essentially drawing the cyclone’s path in the language of graph topology. Even more intriguing is the temporal behavior: across the lifecycle of a storm, from genesis through maturity to decay, the network metrics follow a U-shaped variation, dipping as the cyclone intensifies and recovering as it weakens. This means the network structure is not merely detecting the presence of a closed circulation but is genuinely tracking the storm’s evolution, encoding the degree to which the cyclone has reorganized the pressure field around it.

Detecting a signal, however, is only half the battle; the next challenge was turning it into a reliable classifier. The researchers trained a convolutional neural network—a deep learning architecture originally developed for image recognition—on spatial maps of the four network metrics, teaching it to distinguish ten-day windows containing a tropical cyclone from those that do not. Convolutional networks are well suited to this task because they excel at recognizing local spatial patterns and their arrangements, exactly the kind of track-aligned structures that the network metrics exhibit. The approach follows a lineage of earlier work, including a 2021 study in the same journal by Gupta, Boers, Pappenberger, and Kurths that first demonstrated complex network methods could detect tropical cyclones, but the new framework extends the concept with evolving networks, multiple topological metrics, and a modern interpretability layer.

That interpretability layer may prove to be the study’s most consequential contribution. Deep learning models in the geosciences are often criticized as black boxes: they may achieve impressive skill scores, but scientists cannot tell which features of the input drove a given prediction, making it difficult to trust the model or learn physics from it. To open the box, the team computed saliency maps, a technique that highlights which parts of the input most influence the network’s classification decision. The maps converged on a geographically coherent area spanning the South China Sea and the Philippines, which the authors designate the SCPKR—the South China Sea–Philippines Key Region. When the network topology in this region changes in characteristic ways, the classifier becomes confident that a tropical cyclone is present or imminent in the basin.

Why should this particular corner of the Western North Pacific carry so much informational weight? The answer, the researchers show, lies in the ocean–atmosphere environment that precedes and accompanies storm formation. Composite analysis comparing tropical cyclone windows with non-cyclone windows revealed that the SCPKR exhibits systematically different underlying conditions in the two cases. Two factors stand out. The first is the Western Pacific Subtropical High, the vast semi-permanent anticyclone that dominates the region’s summer circulation; its position and strength modulate steering flows, moisture transport, and the monsoon trough where many cyclones are born. The second is convective activity, the deep cumulonimbus heating that supplies the energy cyclones need to spin up. Variations in the subtropical high and in convection shape how the local pressure field is organized, and that organization is precisely what the network metrics measure. In other words, the topological signature is not an artifact of the method—it is a genuine reflection of the physical environment in which tropical cyclones form.

The implications reach well beyond the Western North Pacific. Tropical cyclone frequency and its response to climate change remain among the most contested questions in climate science, with studies debating whether warming oceans will produce more storms, fewer but stronger storms, or a poleward expansion of the cyclone zone. Any attempt to answer these questions from historical records or model simulations depends on detection methods that are consistent, physically grounded, and robust to the quirks of individual datasets. A network-based detector offers an attractive complement to conventional trackers: it does not depend on a single threshold that may behave differently across reanalysis products and model resolutions, and because it responds to the reorganization of the pressure field as a whole, it may capture forming and weak systems that threshold-based schemes miss. The framework could also be adapted to other basins—the North Atlantic, the Indian Ocean, the South Pacific—where the same physics of pressure reorganization operates, and potentially to other phenomena that restructure atmospheric fields, from atmospheric rivers to blocking events.

There are, of course, caveats and open questions. The study relies on reanalysis pressure data, and the performance of the classifier in operational forecasting settings—where data arrive in real time and are subject to observational gaps—remains to be demonstrated. The ten-day window length, chosen to capture the relevant timescales of cyclone influence on the pressure field, may need tuning for different applications. And while the saliency analysis identifies the SCPKR as the key region, fully exploiting that insight for seasonal prediction or early warning will require linking the network-based detection to conventional forecast skill. Still, the conceptual advance is clear: the atmosphere’s connectivity structure is not just a curiosity of complex systems theory but a practical, information-rich representation of extreme weather. By teaching a machine to read the shape of the climate network, the researchers have shown that tropical cyclones announce themselves not only in wind and rain, but in the very architecture of the pressure field—long before, and in ways, traditional detection methods can see.

Subject of Research: Detection of tropical cyclone occurrence over the Western North Pacific using topological signatures of evolving climate networks constructed from mean sea level pressure anomalies

Article Title: Topological signatures in undirected climate networks enable detection of tropical cyclone occurrence over the Western North Pacific

Article References: Jiang, Z., Wang, M., Li, K., Kurths, J., & Liu, K. (2026). Topological signatures in undirected climate networks enable detection of tropical cyclone occurrence over the Western North Pacific. Climate Dynamics, 64(11), Article 443. https://doi.org/10.1007/s00382-026-08401-y

Image Credits: AI Generated

DOI: 10.1007/s00382-026-08401-y

Keywords: tropical cyclones, climate networks, complex networks, Western North Pacific, mean sea level pressure, convolutional neural network, saliency maps, Western Pacific Subtropical High, South China Sea, network topology, machine learning, Climate Dynamics

The study, led by doctoral researcher Dane Williamson together with Yangfeng Ji and Matthew Dwyer, begins with a deceptively simple observation. When the popular Grad-CAM method is asked to explain why a network chose its top prediction for an image, and then asked again to explain the second-ranked prediction, the highlighted regions barely change. This happens even when the two candidate labels describe semantically unrelated objects, such as an airplane and a barbell. If an explanation method highlights the same pixels whether the model claims to see a bird or an airplane, the researchers argue, that explanation is not telling anyone what distinguishes one hypothesis from another. It is, in effect, a plausible-looking picture with little diagnostic content.

The stakes are higher than they might seem. Prior human-centered research cited in the paper, including the HIVE evaluation study, found that visual explanations can increase user trust even when the underlying model is wrong. In other words, a beautiful heatmap can manufacture confidence in a decision that the heatmap does not actually illuminate. The problem becomes acute in domains such as medical imaging, where documented failures have shown networks latching onto dataset artifacts like text tokens or positional cues instead of actual pathology. If the saliency maps generated in those settings are insensitive to the class label, clinicians auditing a diagnosis could be misled by an explanation that looks informative but is essentially identical to the explanation the model would have produced for any other condition.

To move beyond anecdotes, the researchers formalized a diagnostic test for what they call class sensitivity. For each validation image, they generated saliency maps for the top-1 and top-2 predicted classes and measured the overlap between the top 5 percent of pixels highlighted in each map. A method that genuinely distinguishes between competing labels should produce substantially different maps. They then applied a one-sided Wilcoxon signed-rank test, asking whether the median agreement between maps fell below 50 percent. If a method cannot statistically beat a coin-flip level of overlap, it fails the test of distinguishing one class explanation from another. This framework is deliberately method-agnostic and model-agnostic, meaning it can be applied to any saliency technique and any convolutional backbone.

The results of that diagnostic were sobering. Across five widely used class activation map methods, Grad-CAM, Grad-CAM++, ScoreCAM, AblationCAM, and LayerCAM, evaluated on four architectures spanning ResNet-50, VGG-19, DenseNet-201, and ConvNeXt-Large, and on two datasets, ImageNet and CIFAR-100, most baseline methods failed to distinguish between competing classes in a large fraction of settings. Grad-CAM and AblationCAM in particular frequently produced explanations for the top and second predictions that were statistically indistinguishable. The failure was not confined to one model family or benchmark, suggesting a systemic weakness in how activation-based attribution is computed rather than an artifact of any single network.

Motivated by these findings, the team introduced CASE, short for Contrastive Activation for Class-Sensitive Explanations. The method works by explicitly removing the attribution that a target class shares with its most confusable competitors. Concretely, CASE first computes the gradient of the target class score with respect to the final convolutional activation maps. It then identifies a contrast set consisting of the classes the model most frequently confuses with the target, using the model’s confusion matrix computed on a held-out validation set. The gradients for those contrast classes are averaged into a single vector, and the component of the target gradient that points along this shared direction is subtracted away through an orthogonal projection. What remains is a residual gradient encoding only evidence that is uniquely discriminative for the target class, which is then used to weight the activation maps and produce the final heatmap after a ReLU and upsampling step.

The design choices behind CASE are notable for what they avoid. Unlike prior contrastive explanation approaches such as CWOX, which relies on learned cluster structures, or Contrastive LRP, which is tied to backpropagation-based relevance propagation, CASE requires no architectural modifications, no external baselines, and no auxiliary models. It operates directly on the internal activations of any already-trained convolutional network. The confusion-based contrast selection is also behaviorally grounded: because the contrast set is drawn from classes the model empirically mixes up, the method suppresses exactly the attribution directions that compete with the target in practice, rather than those suggested by semantic similarity alone.

In evaluations across all four architectures and both datasets, CASE was the only method that consistently rejected the null hypothesis of the class-sensitivity test with high confidence, achieving p-values below 0.0001 in every tested setting. ScoreCAM also performed well on most models but stumbled on VGG, while traditional methods failed more often. The study also uncovered a surprising architectural insight: DenseNet proved the most naturally class-sensitive network, and a channel-activity analysis revealed why. Its final convolutional layer activates fewer than 20 channels per class on average, an unusually sparse but highly selective representation, whereas VGG and ConvNeXt activate hundreds of channels that dilute class-specific signals. When the researchers swapped DenseNet’s attribution layer to the final normalization layer, which activates nearly all channels uniformly, most methods lost their class sensitivity entirely. Even retraining across five random seeds confirmed the effect was a stable property of the architecture, and CASE retained its sensitivity even in the diffuse regime.

Class sensitivity alone is not sufficient, of course. A good explanation should also be faithful, meaning the highlighted regions should genuinely matter to the model’s output. The team tested this by ablating the top 5 percent of salient pixels and measuring the drop in the model’s confidence. Here the picture was more nuanced. CASE matched the strongest baselines on DenseNet, ResNet, and ConvNeXt, and significantly outperformed AblationCAM and LayerCAM on ConvNeXt, but on VGG three methods, Grad-CAM++, LayerCAM, and ScoreCAM, produced larger confidence drops. The authors are candid about these failure cases, noting that uniquely discriminative regions are not always the regions whose removal maximally reduces confidence. When class evidence is diffuse and shared across labels, perturbing broadly activated regions can produce bigger confidence swings even if those regions are less informative for separating one class from another.

The paper also probes the robustness of the method itself. A pre-registered degenerate-residual check searched for cases where the target and contrast gradients were nearly collinear, which would collapse the orthogonal projection toward zero; across 4,774 generated maps, not a single near-zero residual was found. Ablations over the contrast set size showed that three confused classes are sufficient, with larger sets yielding no consistent benefit. The authors are explicit about the method’s boundaries: CASE applies to single-label classification with convolutional networks, and extending it to transformers, multi-label tasks, or multimodal models would require rethinking the contrast formulation. Those limitations aside, the study delivers a message that should resonate far beyond the machine learning community. Explanation tools are trusted as safeguards, and this work shows that trust must be earned through formal diagnostics rather than visual plausibility. As deep learning spreads into medicine, safety monitoring, and high-stakes auditing, the researchers argue, we need explanations that reveal what truly sets a model’s decision apart, not just what makes a convincing picture.

Subject of Research: Class sensitivity and contrastive saliency methods for explaining convolutional neural network image classifiers

Article Title: CASE: Contrastive Activation for Class-Sensitive Explanations

Article References: Williamson, D., Ji, Y., & Dwyer, M. (2026). CASE: Contrastive Activation for Class-Sensitive Explanations. Machine Learning, 115(10), Article 230. https://doi.org/10.1007/s10994-026-07169-w

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07169-w

Keywords: explainable AI, saliency maps, Grad-CAM, contrastive explanations, convolutional neural networks, class activation mapping, machine learning, attribution fidelity, ImageNet, CIFAR-100, DenseNet, model interpretability