PaperlayerView exact revision on arXiv

Source-linked AI summary

Component Pooling Trained Across Orientations for Few-Shot Remote Sensing Scene Classification

Bingchen Yan, Wenheng Guo, Yuye Zhou

arXiv:2610.33518v1cs.CV

TL;DR

FSRSSC must recognize novel scene categories from few labels, while common pooling treats scene-defining objects and background alike and aerial images lack a canonical orientation. RoCoP combines component pooling with rotated-view backbone training. Across benchmarks, the components are complementary, RoCoP outperforms re-implemented methods under the same protocol, and very small base sets limit its usefulness.

  • Problem

    FSRSSC must recognize novel remote-sensing scene categories from few labeled images, while common global average pooling weights object and background positions alike and aerial images lack a canonical orientation.

    Source: §1 · p. 2

  • Method

    RoCoP uses parameter-free NDM and CSM to pool feature components and trains its backbone to classify randomly rotated views of each image.

    Source: §1 · p. 2, §3.5 · p. 7

  • Results

    Across three benchmarks, RoCoP outperforms re-implemented methods under the same protocol and improves other methods as a plug-in; its components together improve the baseline by 2.51 points in 1-shot accuracy.

    Source: §4.5 · p. 10, §6 · p. 15

  • Takeaways & Limitations

    Component pooling and rotated-view training are complementary, and NDM and CSM can improve other methods as plug-in modules.

    Source: §4.5 · p. 10, §6 · p. 15

    Component pooling lowers accuracy with the 1,000 base images of UC Merced, and single-seed training leaves between-run variance outside the reported intervals.

    Source: §5 · pp. 14–15

Abstract

from arXiv · show

Few-shot remote sensing scene classification (FSRSSC) aims to recognize novel scene categories from only a few labeled images. Most FSRSSC methods summarize the feature map of an image by global average pooling, which treats the few objects that define an aerial scene and the large background around them alike, and they rarely consider that aerial images have no canonical orientation. To address these issues, we propose RoCoP, a parameter-free component pooling network trained across orientations. RoCoP consists of a feature extractor, a non-negative decomposition module (NDM) and a component scoring module (CSM). NDM factorizes the local descriptors of each image into a few additive components with a deterministic initialization and a single multiplicative update, so that the positions of one object are grouped together. CSM weights each component by its activation magnitude and converts the weights into a soft mask for pooling. Because both modules are invariant to permutations of feature positions, the only rotation-sensitive part of the pipeline is the backbone, which we train to classify randomly rotated views of every image. Experiments on NWPU-RESISC45, AID and UC Merced show that component pooling and rotated-view training are complementary: together they improve a strong transfer-learning baseline by about 2.5 points in 5-way 1-shot accuracy on NWPU-RESISC45 and AID, without adding parameters at inference. Under an identical protocol, RoCoP outperforms re-implemented metric-based and task-adaptive methods, and NDM and CSM also improve existing methods when plugged into them. We further identify the regime in which the method does not apply, namely very small base sets.

1. Introduction

Few-shot remote sensing scene classification asks models to recognize novel categories from very few labeled images, despite aerial scenes’ object–background structure and lack of canonical orientation. RoCoP addresses these challenges by pooling feature components and training its backbone across rotated views.

1. Introduction

  • FSRSSC trains on base classes and recognizes novel scene categories from one or five labeled images per class.Remote sensing scene classification assigns semantic labels to aerial or satellite images.Source: §1 · pp. 1–2
  • Global average pooling weights object and background positions equally, while aerial images have no canonical orientation.Scene evidence often comes from a few objects surrounded by much larger areas of bare land, vegetation, or water.Source: §1 · p. 2
  • RoCoP pools over components rather than positions and trains its backbone across orientations using randomly rotated views.NDM groups positions with alike channel responses, and CSM weights each component as a unit; neither module has learnable parameters.Source: §1 · p. 2
  • RoCoP outperforms re-implemented methods, improves other methods as a plug-in, and is limited on very small base sets.The reported analysis also finds that plain rotated-view training is preferable to explicit consistency constraints.Source: §1 · p. 2

2. Related Work

Related work spans few-shot learning paradigms, remote-sensing adaptations, feature pooling, and rotation handling. RoCoP retains transfer learning while changing feature pooling and training views.

2.1. Few-Shot Learning

  • Transfer-learning methods train an embedding on base classes, then fit a simple classifier on each task’s support set; this paper adopts that setting.The section contrasts this approach with optimization-based and metric-based few-shot learning.Source: §2.1 · p. 3

2.2. Few-Shot Remote Sensing Scene Classification

  • Remote-sensing few-shot work includes attention, prototype calibration, metric and subspace methods, contrastive learning, relation modeling, and multi-scale fusion.RoCoP keeps the standard transfer-learning pipeline and changes pooling and training views, allowing combination with other methods.Source: §2.2 · p. 3

2.3. Feature Weighting and Pooling

  • Prior pooling and attention approaches weight positions independently or learn their weights on base classes; NDM groups positions first using NMF.The overview depicts NDM components and CSM scores forming a soft pooling mask.Source: §2.3 · p. 3, Figure 1 · p. 4

2.4. Rotation in Remote Sensing

  • Rotation-aware networks encode orientation through specialized layers or self-supervised rotation prediction; RoCoP instead classifies rotated copies with the original label.The paper reports that this training changes what component pooling learns.Source: §2.4 · p. 4

3. Method

RoCoP pools image features over components formed from local descriptors, scores those components to create a soft mask, and trains its backbone on rotated views. Its parameter-free pooling modules are permutation-invariant, while the backbone remains rotation-sensitive.

3.1. Problem Setting and Overview

  • RoCoP evaluates on disjoint base and novel classes: a frozen ResNet-12 backbone supplies features for inductive few-shot episodes.Queries are classified independently without sharing statistics across the query set.Source: §3.1 · p. 4

3.2. Non-negative Decomposition Module (NDM)

  • NDM groups similar non-negative local descriptors into additive components, allowing CSM to weight positions belonging to one object as a unit.A deterministic initialization and one multiplicative update produce the decomposition for each image.Source: §3.2 · p. 5

3.3. Component Scoring Module (CSM)

  • CSM scores each component using assignment-weighted descriptor activation magnitudes, calibrates scores per image, and transfers component weights to positions as a soft mask.The intended weighting gives stronger, specific responses higher weights and flat background lower weights; multiple components can contribute to the mask.Source: §3.3 · p. 6

3.4. Component Pooling and Classification

  • The mask pools descriptors for classification; RoCoP includes GAP when there is one component, while more components allow weighting distinct descriptor groups.The position-wise control uses descriptor norms without grouping, isolating the effect of the grouping step.Source: §3.4 · pp. 6–7

3.5. Rotated-View Training

  • NDM and CSM are invariant to feature-position permutations, so the backbone is the remaining source of rotation sensitivity and is trained on randomly rotated views with the original labels.At test time, images are not rotated; an additional mask-consistency loss was tested but lowered accuracy and is excluded from RoCoP.Source: §3.4 · p. 7, §3.5 · p. 7

3.6. Complexity

  • NDM and CSM add no parameters, leaving the backbone at 12.42 M parameters; RoCoP pooling takes 67 ms for a 5-way 5-shot episode versus 44 ms with GAP.The reported timings use one RTX 4090 and an episode with 25 support and 75 query images.Source: §3.6 · p. 8

4. Experiments

Experiments use NWPU-RESISC45, AID, and UC Merced with separate base, validation, and novel classes. Models are trained on base classes and evaluated on shared few-shot episodes for consistent comparisons.

4.1. Datasets

  • NWPU-RESISC45, AID, and UC Merced provide 45, 30, and 21 scene classes, respectively, with separate base, validation, and novel class splits.Their splits contain 25/10/10, 16/7/7, and 10/5/6 classes, respectively; UC Merced has 1,000 images for base training.Source: §4.1 · p. 8

4.2. Implementation Details

  • A ResNet-12 is trained from scratch on base classes for 100 epochs, with five NDM components, one multiplicative update, and validation-selected checkpoints.The training setup uses SGD and evaluates checkpoints on 300 validation episodes every five epochs; novel classes are reserved for final evaluation.Source: §4.2 · p. 8
  • Mean 5-way 1-shot and 5-way 5-shot accuracy is measured over 2,000 shared test episodes, and configurations isolate component pooling, rotated-view training, and mask consistency.Each episode has 15 queries per class; paired 95% confidence intervals reflect episode variability, not variation across training runs.Source: §4.2 · p. 8

4.3. Comparison with Published Methods

Under the same protocol, RoCoP outperforms re-implemented methods on both datasets, while plug-in studies find that NDM and CSM improve two host methods. Comparisons with published results are qualified by differences in settings.

4.3. Comparison with Published Methods

  • RoCoP improves GAP on NWPU-RESISC45 and AID and beats re-implemented methods, but trails the strongest published NWPU methods and its own baseline on UC Merced.Published settings differ, so Table 1 is a reference rather than a ranking.Source: §4.3 · pp. 8–9, Table 1 · p. 9

4.4. Same-Protocol Comparison and Plug-in Study

  • The fair comparison re-implements ProtoNet, FEAT, FRN and DeepEMD with a shared backbone, input size, class splits, checkpoint selection and test episodes.Each method retains its own training objective.Source: §4.4 · p. 9
  • Adding NDM and CSM improves ProtoNet and rotation prediction on both datasets; the latter reaches 72.18% 1-shot accuracy on NWPU-RESISC45, 2.32 points above its host.On AID, the rotation-prediction plug-in gains 0.47 points but remains below RoCoP; rotation prediction alone does not improve the rotated-view GAP model.Source: §4.4 · p. 9
  • RoCoP achieves the highest accuracy among the compared methods on both datasets, while FEAT is 1.32 points lower (1-shot) on NWPU-RESISC45 and 1.87 points lower on AID.Source: §4.4 · p. 9

4.5. Ablation Study

The ablations find that component pooling and rotated-view training work complementarily, while explicit mask consistency reduces accuracy. Results vary by scene structure, and the classifier choice does not change the main conclusions.

Mask-consistency loss

  • Adding explicit mask-consistency loss lowers accuracy; rotated-view training already makes masks more consistent, and further constraints appear to remove useful flexibility.Source: §4.5 · p. 10

Test classifier

  • The mask-weighted prototype classifier yields about one point less accuracy than logistic regression, but the conclusions hold and RoCoP's gains come from features rather than classifier choice.Source: §4.5 · p. 10

Effect of each part

  • 2.51 points (1-shot) and 1.73 points (5-shot): component pooling plus rotated-view training improve GAP, with positive interactions across shot counts and classifiers.Component pooling's 1-shot gain roughly doubles when used with rotated views.Source: §4.5 · p. 10, Table 4 · p. 11

Per-class analysis

  • RoCoP gains most on structured-object scenes, while texture or repeated-unit scenes lose accuracy, consistent with grouping helping where distinct objects lie against backgrounds.Source: §4.5 · pp. 10–11, Figure 2 · p. 11

Calibration

  • Table 5 compares per-image and support-level weight calibration on NWPU-RESISC45 using the same checkpoints and 2000 episodes.Source: Table 5 · p. 12

4.6. Per-Image versus Support-Level Calibration

Per-image calibration is retained because support-level calibration lowers 1-shot accuracy for component models. Across datasets, RoCoP improves GAP on AID, while results also show a small-base-set boundary and changes in pooling-mask behavior under rotation.

4.6. Per-Image versus Support-Level Calibration

  • Support-level calibration consistently lowers component models’ 1-shot accuracy and leaves their 5-shot accuracy unchanged, so RoCoP retains per-image calibration.Support-level calibration slightly improves position-wise weighting; the authors suggest it compresses component weights relative to the per-image calibration used in backbone training.Source: §4.6 · p. 12

4.7. Results Across Datasets

  • RoCoP improves GAP by 2.53 points in 1-shot accuracy on AID, where component grouping also outperforms position-wise weighting at both shot counts.The two parts interact positively, and the same pattern holds with the prototype classifier.Source: §4.7 · p. 12, Table 6 · p. 13
  • With 1,000 UC Merced base images, component pooling lowers accuracy with and without rotated views, leaving RoCoP below GAP for both classifiers.Rotated-view training improves the component model but does not offset the loss; the authors attribute the limitation to too few images for learning reliable object-background descriptors.Source: §4.7 · p. 12

4.8. Rotation Behaviour and Visualization

  • On novel NWPU-RESISC45 classes, models are almost invariant to 90° query rotations, but RoCoP’s masks agree better with back-rotated original masks than masks from NDM without rotated-view training.The mask shift occurs across the test-image distribution, not only in its mean.Source: §4.8 · p. 12
  • NDM pooled vectors are slightly less rotation-invariant than GAP vectors because the mask depends on positions while GAP ignores them.Qualitative mask examples are provided in Appendix A.Source: §4.8 · p. 13

4.9. Number of Components

The selected component count, r = 5, gives the best accuracy among the tested settings, while accuracy is stable across r = 3 to r = 8. Smaller counts can make training less stable.

4.9. Number of Components

  • r = 5 gives the best accuracy among tested component counts, and accuracy changes by less than half a point from r = 3 to r = 8.Table 8 reports paired differences over the same 2,000 episodes on NWPU-RESISC45 test classes.Source: §4.9 · p. 13, Table 8 · p. 14
  • A first run with r = 3 diverged after epoch 40, although a repeated run trained normally.The authors do not tune the number of decomposition updates in this paper.Source: §4.9 · p. 13
  • With few components, per-image calibration can assign weights of 0 and 1 to different components, switching large image regions off or on together and potentially destabilizing optimization.Source: §4.9 · p. 13

5. Discussion

The discussion identifies the settings where component pooling helps and where its evidence is limited, including small base sets, dataset-dependent grouping gains, and constraints on calibration and evaluation.

5. Discussion

  • Component pooling lowers accuracy on UC Merced’s 1,000-image base set, while the authors state it is suited to base sets of several thousand images.Source: §5 · p. 14
  • On NWPU-RESISC45, norm-based position weighting already captures much of reweighting’s benefit and grouping adds a clear gain only in 5-shot; on AID, grouping helps at both shot counts.Source: §5 · p. 14
  • Support-set calibration did not help with a backbone trained using per-image calibration, and single-seed training means reported intervals omit between-run variance.The GAP baseline is below the best published NWPU-RESISC45 methods, so the results establish a relative gain under one protocol rather than a new state of the art.Source: §5 · p. 15

6. Conclusion

The conclusion presents RoCoP as parameter-free component pooling trained across orientations for few-shot remote sensing scene classification. It reports gains over re-implemented methods under the same protocol and improvement of other methods when RoCoP’s modules are used as plug-ins.

6. Conclusion

  • RoCoP combines NDM, which groups local descriptors into additive components, with CSM, which weights components into a soft pooling mask, and trains the backbone with rotated views.Source: §6 · p. 15
  • Across three benchmarks, component pooling and rotated-view training are complementary, and RoCoP outperforms re-implemented methods under the same protocol.Source: §6 · p. 15
  • NDM and CSM can improve other methods when used as a plug-in module.Source: §6 · p. 15

Related papers

Loading 2610.33518v1…