In this paper we evaluate the performance of topological features for generalizable and explainable classification of firn image data, with the broader goal of understanding the advantages, pitfalls, and trade-offs in topological featurization. Firn refers to layers of granular snow within glaciers that haven't yet been compressed into ice. This compaction process imposes distinct topological and geometric structure on firn that varies with depth within the firn column. However, the limited and highly correlated training data available results in naive classifiers (such as convolutional neural networks) overfitting to orientation-dependent features. This motivates us to evaluate topological data analysis (TDA) as a way to design features which capture the shape of the firn data while retaining a large degree of invariance. We evaluate two classes of topological features, sublevel set features and distance transform features, together with persistence curves, to predict sample depth from firn images. A range of challenging training-test scenarios reveals that no one choice of method dominates in all categories. In particular, we see a tradeoff between accuracy on the one hand, and generalizability and explainability (via feature importance) on the other.
Figures & tables
Figure 1: One sample image per depth for the firn dataset. Gray areas represent ice, while the black areas show the pore (air) space.
Figure 2: Computing the 0\textsuperscriptth Betti curve (b) from the 0\textsuperscriptth -dimensional persistence diagram (a). The red dashed lines indicate the extent of the fundamental box at each threshold t , with these thresholds marked with a red dot on the Betti curve.
Figure 3: The first 8 samples from the 23m depth, showing a high degree of correlation in the overall shape of the ice and pore space.
Figure 4: Simple CNN architecture to evaluate overfitting on the firn dataset when no topological features are used. ReLU was used as the activation function for the hidden layers, and softmax for the output layer. We trained for a maximum of 50 epochs with batch size 20 and early stopping using a 20% validation split, Adam as the optimizer, and with cross-entropy as our loss function.
Category prediction (% accuracy)
Unaltered test set
99.13±0.20
Test set with Gaussian blur
99.07±0.22
Test set with random symmetries applied
31.68±1.05
Table 1: Results of the CNN on different test sets. We show the mean and standard error over ten trials.
Figure 5: Visual representations of the two topological featurizations for an example image, using Betti curves as an example. In each sequence, blue indicates an included pixel in the cubical complex.
Figure 6: The original image and two manipulations used to evaluate predictor generalizability. The original image is shown in (a). In (b), we show the quadrants used for the Split and Split BR scenarios. The blurred version of the image (applied to the test set in Blurred ) is shown in (c).
Scalar prediction (mean absolute error in metres)
vBettiSS
vGaussianSS
vBettiDT
vGaussianDT
Whole
0.74 ± 0.12
0.90 ± 0.14
4.01 ± 0.29
3.86 ± 0.25
Split
2.10 ± 0.10
1.92 ± 0.10
7.74 ± 0.15
8.36 ± 0.18
Split BR
2.40 ± 0.04
2.21 ± 0.05
7.35 ± 0.02
7.86 ± 0.03
Blurred
14.61 ± 0.78
11.26 ± 0.93
3.97 ± 0.27
3.97 ± 0.25
Missing depths
9.93 ± 0.22
10.15 ± 0.26
6.20 ± 0.04
7.69 ± 0.10
Table 2: Performance measures for topological feature-based depth prediction. When predicting depth as a scalar value we use mean absolute error (MAE). Note that true values range from 7 m to 78 m. For predicting depth as one of ten predefined levels (see Figure 1 ), we use accuracy. Ten trials are performed to obtain each entry, with the mean and standard deviation shown.
Figure 7: On the left, the average Betti curve vector vBettiSS by depth for the sublevel set featurization, showing good separation but conspicuously high peak values. On the right, we show a sample image from the 78 m depth threholded at t=120 , near the peak of the average curve for that depth. We see that the topology of the image at this threshold is dominated by granular noise.
Figure 8: Examining two important features for the sublevel set featurization. We show sample images thresholded at the relevant thresholds with their corresponding Betti numbers (above). We also show a labeled scatter plot using the two features (below).
Figure 9: Examining two important features for the distance transform featurization. We show the distance transform for sample images thresholded at the relevant thresholds with their corresponding Betti numbers (top). We show a labeled scatter plot using the two features (middle), with tiny amounts of noise added so that multiple points at integer coordinates are visible. Boxplots of feature values by depth are also shown for each feature (bottom).
Recent advances in pretraining 3D point cloud encoders (e.g., Point-BERT, Point-MAE) have produced powerful models, whose abilities are typically evaluated on geometric or semantic tasks. At the same time, topological descriptors have been shown to provide informative summaries of a shape's multiscale structure. In this paper we pose the question whether topological information can be derived from features produced by 3D encoders. To address this question, we first introduce DONUT, a synthetic benchmark with controlled topological complexity, and propose FILTR (Filtration Transformer), a learnable framework to predict persistence diagrams directly from frozen encoders. FILTR adapts a transformer decoder to treat diagram generation as a set prediction task. Our analysis on DONUT reveals that existing encoders retain only limited global topological signals, yet FILTR successfully leverages information produced by these encoders to approximate persistence diagrams. Our approach enables, for the first time, data-driven extraction of persistence diagrams from raw point clouds through an efficient learnable feed-forward mechanism.
Building practical filtrations on objects to detect topological and geometric features is an important task in the field of Topological Data Analysis (TDA). In this paper, leveraging the ability of the Laplacian of Gaussian operator to enhance the boundaries of medical images, we define the G-LoG (Gaussian-Laplacian of Gaussian) bi-filtration to generate the features more suitable for multi-parameter persistence module. By modeling volumetric images as bounded functions, then we prove the interleaving distance on the persistence modules obtained from our bi-filtrations on the bounded functions is stable with respect to the maximum norm of the bounded functions. Finally, we conduct experiments on the MedMNIST dataset, comparing our bi-filtration against single-parameter filtration and the established deep learning baselines, including Google AutoML Vision, ResNet, AutoKeras and auto-sklearn. Experiments results demonstrate that our bi-filtration significantly outperforms single-parameter filtration. Notably, a simple Multi-Layer Perceptron (MLP) trained on the topological features generated by our bi-filtration achieves performance comparable to complex deep learning models trained on the original dataset.
Recent advances in attention-based deep learning have motivated their adoption for remote sensing image classification; however, their benefits for cryospheric imagery, where surface states are dominated by fine-grained textures and class imbalance, remain unclear. In this work, we revisit a benchmark Greenland Ice Sheet image dataset, previously shown to favor convolutional neural networks (CNNs), to examine whether modern attention-based and hybrid architectures improve class-wise reliability. We conduct a controlled comparison between a classical CNN (AlexNet), a modern CNN (ConvNeXt-Tiny), a pure attention-based model (Swin-Tiny), and a hybrid convolution-attention model (CoAtNet-0) under identical training and evaluation protocols. Results show that AlexNet achieves the highest accuracy and the strongest balanced performance as measured by macro-averaged F1, while ConvNeXt-Tiny exhibits the highest macro-averaged AUC, indicating strong class separability but less consistent final decision quality. Class-wise analysis reveals that hybrid architectures improve recall for rare and structurally distinct surface classes, whereas convolutional models remain more reliable for texture-dominated categories. These findings highlight the importance of aligning architectural inductive bias with cryospheric data characteristics and suggest that increased model complexity does not necessarily translate to improved reliability for ice-sheet surface classification.
Chhaya Kulkarni, Emam Hossain
Department of Computer and Information Sciences Towson University · Department of Information Systems University of Maryland, Baltimore County