PaperlayerView exact revision on arXiv

Source-linked AI summary

Recent Advances in Domain Adaptation for the Classification of Remote Sensing Data

Devis Tuia, Claudio Persello, Lorenzo Bruzzone

arXiv:2104.07778v1cs.CVeess.IV

TL;DR

Domain adaptation addresses remote-sensing settings where data distributions are not stationary over time or space, while validation remains a bottleneck for employing adaptation strategies. The paper reviews recent approaches across four families, reports improved target-domain generalization and compares active target querying with random queries, then offers selection guidelines for real applications.

  • Problem

    Nonstationarity across domains and the validation of adaptation strategies constrain the use of domain adaptation for remote sensing classification.

  • Method

    The paper reviews recent domain-adaptation models grouped into invariant-feature selection, data-representation matching, classifier adaptation, and selective sampling.

  • Results

    91.0% overall accuracy was obtained on the source domain and 80.7% on the target domain for one feature set, while another result reports improved generalization across domains under dataset shift.

  • Takeaways & Limitations

    The review provides guidelines for selecting domain-adaptation approaches in real application scenarios and identifies domain adaptation as important for future Earth Observation.

Abstract

from arXiv · show

The success of supervised classification of remotely sensed images acquired over large geographical areas or at short time intervals strongly depends on the representativity of the samples used to train the classification algorithm and to define the model. When training samples are collected from an image (or a spatial region) different from the one used for mapping, spectral shifts between the two distributions are likely to make the model fail. Such shifts are generally due to differences in acquisition and atmospheric conditions or to changes in the nature of the object observed. In order to design classification methods that are robust to data-set shifts, recent remote sensing literature has considered solutions based on domain adaptation (DA) approaches. Inspired by machine learning literature, several DA methods have been proposed to solve specific problems in remote sensing data classification. This paper provides a critical review of the recent advances in DA for remote sensing and presents an overview of methods divided into four categories: i) invariant feature selection; ii) representation matching; iii) adaptation of classifiers and iv) selective sampling. We provide an overview of recent methodologies, as well as examples of application of the considered techniques to real remote sensing images characterized by very high spatial and spectral resolution. Finally, we propose guidelines to the selection of the method to use in real application scenarios.

I. INTRODUCTION

Remote-sensing classification must handle acquisition, temporal, spatial, and sensor-related shifts while often relying on labels from only one image. The paper presents domain adaptation as a framework for reusing models across related but different data distributions and organizes recent methods into four families.

  • Motivation: Supervised models often require image-specific reference samples, but extensive labeling cannot keep pace with frequent acquisitions, large areas, or costly ground campaigns.The challenge is especially acute for quantities such as chlorophyll concentration, plant water stress, and tree species that cannot always be obtained through photointerpretation.
  • Motivation: Applying a pretrained model directly to a new dataset often performs poorly because corresponding object spectra differ across scenes.The cited shifts can arise from nonrepresentative spatial sampling, illumination or acquisition-angle changes, and seasonal variation.
  • Challenges: Models trained on one sensor are difficult to reuse with another because many classifiers require equal dimensionality and matching meanings for every input feature.Existing fusion strategies generally cover only certain sensor combinations and commonly use shared bands, limiting model reuse and multisensor synergies.
  • Requirements: Remote-sensing processing systems therefore need robustness to changing acquisition conditions, temporal shifts, and ideally sensor differences.These requirements are linked to applications involving multitemporal, multisource, and multiresolution imagery, including post-catastrophe interventions and large-scale monitoring.
  • Domain adaptation: Domain adaptation adapts a model trained for a specific task to a new but related task when the original model provides useful, though imperfect, knowledge.The paper connects this framework to remote-sensing classification across altered data distributions and related multisource settings.
  • Paper scope: The review introduces recent remote-sensing DA methods, evaluates their strengths and weaknesses through practical high-resolution examples, and proposes application-oriented selection guidelines.Its taxonomy comprises invariant feature selection, representation matching, classifier adaptation, and limited but effective sampling in new domains.

II. TRANSFER LEARNING AND DOMAIN ADAPTATION

Domain adaptation addresses classification when source and target remote-sensing domains differ but remain related enough for source information to help. The section distinguishes supervised, semisupervised, and unsupervised settings by the label information available and formalizes adaptation between joint or marginal distributions.

  • Domain adaptation transfers a classifier trained on a source domain to make predictions on a target domain with a different joint distribution.Source and target images may cover different areas or the same area at different times; differences can arise from illumination, viewing angle, soil moisture, or topography.
  • Supervised DA uses labeled samples from both domains, typically with far fewer target labels than source labels.The setting assumes m << n and therefore emphasizes effective use of source-domain information.
  • Semisupervised DA uses labeled source samples and only unlabeled target samples, making it more challenging and dependent on assumptions linking the domains.The method family requires sufficient similarity between source and target distributions for adaptation to converge consistently on the target.
  • All DA methods assume source and target distributions differ but are close enough for source information to benefit target learning.If the domains are arbitrarily different, source information offers no expected advantage; if their joint distributions are equal, adaptation is unnecessary.
  • Unsupervised DA has no labels in either domain and matches source and target marginal distributions without task-specific label information.These methods can serve as preprocessing for classification, clustering, or density estimation, but require similar structural properties across datasets.

adaptation. Unsupervised DA models are generally feature extractors or matching algorithms

Remote-sensing transfer problems combine domain shifts with biased or nonrepresentative training samples. The section contrasts general domain adaptation with sample selection bias and covariate shift, while noting that these effects can be entangled in real applications.

  • Training samples are often selected and labeled through photo-interpretation or field surveys, making sampling bias plausible in remote-sensing applications.The reviewed techniques target transfer-learning problems regardless of the cause of the dataset shift.
  • Covariate shift is a special sample-selection-bias case in which bias depends only on inputs X, not labels Y.Its defining approximation is P_s(Y|X) ≈ P_t(Y|X), while P_s(X) ≠ P_t(X).
  • Domain adaptation can change class-conditional densities across domains, creating a substantially different classification problem and potentially misleading source training samples.Target classes may overlap different source classes, making direct transfer unreliable.
  • Sample selection bias draws source and target samples with bias from the same underlying distribution, producing a milder shift than general domain adaptation.Cross-domain class overlap is not likely in this case.
  • Real remote-sensing problems may simultaneously contain acquisition-condition differences and nonrepresentative training samples.These effects can bias class-probability estimates and produce inaccurate classification models.

III. A TAXONOMY OF ADAPTATION METHODS

The paper organizes remote-sensing domain-adaptation methods into four families based on whether they alter features, distributions, classifiers, or target sampling. The families differ in their use of unlabeled versus labeled target data.

  • The taxonomy contains invariant-feature selection, representation matching, classifier adaptation, and active-learning-based classifier adaptation.These categories group recent methods according to the component used to address cross-domain shift.
  • Invariant-feature methods identify input features unaffected by shifting factors before classifier training, removing unstable features from the representation.The resulting feature space is intended to be more stable across domains.
  • Representation-matching methods make source and target data distributions similar and extract a common space while leaving the classifier unchanged.They generally operate on the original input spaces through joint feature extraction.
  • Classifier-adaptation methods retain the data distributions and adapt a source-trained model using unlabeled target samples through semisupervised strategies.The adaptation acts on the classifier rather than constructing a shared data representation.
  • Active-learning adaptation iteratively retrains the model with a limited number of carefully selected labeled target samples.Selection matters because acquiring target labels is costly and samples should have strong potential to move the model toward the desired target classifier.

A. Selecting invariant features

Invariant-feature selection seeks subsets that remain discriminative and stable across domains, especially under hyperspectral spectral shifts. A multi-objective strategy identifies trade-offs between relevance and invariance, improving target-domain generalization in the reported example.

  • A. Selecting invariant features: An alternative invariance strategy augments training with synthetic labeled examples representing physically consistent illumination, size, and rotation changes.For SVMs, variations are generated only for source-trained support vectors.
  • A. Selecting invariant features: Hyperspectral imagery is highly sensitive to acquisition changes, which can make class spectral signatures nonstationary across domains.The paper illustrates this issue with Hyperion imagery acquired over two Okavango Delta areas in Botswana.
  • A. Selecting invariant features: Invariant-feature selection jointly evaluates class discrimination and feature stability between source and target domains.The subset F is optimized using relevance term Δ and dataset-shift term P, with subset size l fixed.
  • A. Selecting invariant features: A multi-objective genetic algorithm finds Pareto-optimal feature subsets, avoiding the need to specify a relative weighting between relevance and shift.Solutions represent trade-offs between discriminative power and cross-domain stability.
  • A. Selecting invariant features: Target-domain accuracy is highest for feature subsets balancing relevance and invariance, whereas relevance alone favors source-domain accuracy without sufficient target stability.The results indicate that P captures feature stability and that trade-off solutions support generalization across domains.
  • A. Selecting invariant features: 91.0% source OA and 80.7% target OA were obtained with the selected subset, versus 92.7% source and 64.4% target OA when optimizing Δ alone.Accounting for dataset shift substantially improved the reported target-domain generalization.

B. Adapting data distributions

Representation-matching methods adapt remote-sensing data by aligning source and target representations, using transformations that may accommodate unpaired, differently dimensional, multisource, nonlinear, and partially labeled domains. The reviewed literature includes feature extraction, manifold, kernel, graph, spectral, and distribution-alignment approaches, with method suitability depending on the application constraints.

  • Alignment requirements: Desired alignment properties include handling unpaired data, different dimensionality, multiple domains, nonlinear transformations, and incomplete labeling.These properties target non-coregistered data, multisource classification, multitemporal adaptation, atmospheric or illumination effects, and unavailable labels.
  • Representation matching: Representation matching aligns data spaces so a common classifier or regressor can be applied across domains.Most recent work focuses on feature extraction strategies that align the data spaces.
  • Method families: Reviewed methods include histogram, PCA or kPCA, MMD, CCA, manifold, eigendecomposition, graph-matching, and multispectral transformations.The approaches differ in whether they use statistical distances, geometric structure, semantic ties, spatial information, or spectral correspondences.
  • Method trade-offs: Some methods align arbitrary numbers of differently dimensional domains without paired examples, but require labeled samples in every domain.This illustrates the trade-off between broad alignment capability and labeling requirements.
  • Method selection: Method selection depends on domain structure: canonical-correlation methods require coregistration, whereas non-multiview methods cannot align more than two domains.The paper directs analysts to use its summary table to select an approach for the specific situation.
  • Remote-sensing example: In the Rio angular-sequence experiment, source-only prediction performed poorly at strong off-nadir angles, while SSMA provided particularly good performance.The experiment transfers a classifier from a nadir image to non-coregistered target images acquired in a single WorldView2 pass.

C. Adapting classifiers with semisupervised approaches

Semisupervised classifier adaptation uses unlabeled or sparsely labeled target data to modify models trained on source data. The reviewed methods include mixture-model, tree, SVM, kernel, multitask, and manifold-regularization approaches, generally assuming shared classes and features.

  • Semisupervised adaptation: Semisupervised adaptation uses unlabeled target samples to account for divergence between source and target domains.Kernel-based approaches can select decision functions that both project data discriminatively and minimize training-test shift.
  • Assumptions and extensions: Most reviewed methods assume that source and target domains share the same classes and features, although one approach addresses differing class sets.That approach combines unsupervised change detection with the Jeffreys-Matusita distance to detect appearing or disappearing classes.
  • Distribution-based adaptation: Mixture-model methods update an already trained classifier using the new image distribution, often estimating components with EM and a finite Gaussian mixture model.Related extensions use Bayesian cascades, multiple classifiers, and multiple-cascade architectures.
  • Classifier ensembles: Tree-based adaptation uses target-domain density functions or randomized trees, with a semisupervised extension that downweights poorly performing classifiers.The randomization variant does not control the adaptation objective.
  • SVM adaptation: SVM approaches adapt source decision boundaries toward low-density target regions or regularize the solution using source-target distribution divergence.Other methods share information across domain-specific tasks or regularize labeled and unlabeled samples through graph geometry.
  • Iterative adaptation: DASVM iteratively adds selected unlabeled target samples to the cost function while gradually removing source training samples.At convergence, the method can classify target samples accurately.

D. Adaptation of the classifier by active learning

Active-learning domain adaptation interactively labels informative target samples so a source-trained classifier can evolve toward the target distribution. In the KSC experiment, combining active learning with TrAdaBoost made early queries substantially more effective than random sampling.

  • Active-learning principle: Active learning queries target samples selected by model uncertainty, allowing the classifier to migrate toward an optimal target-domain model.The process starts with a source-optimal classifier and gradually evolves as target labels are supplied.
  • New classes: Active learning can also identify candidate regions for new target classes by screening feature-space clusters or assessing change uncertainty.These strategies present dense unlabeled clusters or uncertain changes to the user for labeling.
  • Handling strong shifts: When source-target deformation is strong, source labels can become harmful because target-class distributions may overlap different source classes.Reweighting or removing contradictory source samples lets the adapted classifier specialize to the target domain.
  • Large-area processing: For very large areas, cluster selection, compressed sensing, and geographically distributed search reduce the number of active-learning iterations.The paper identifies area size as a setting requiring specific active-learning solutions.
  • Weight evolution: During TrAdaBoost plus active learning, target support-vector weights increased while source weights stabilized at 40% of the original training samples.The weighting pattern indicates reduced source importance and strong weight assigned to each newly added target sample.

IV. GUIDELINES FOR THE CHOICE OF THE ADAPTATION STRATEGY

The paper provides guidelines for choosing an adaptation strategy and separately discusses validation of adapted models. Its summary table is intended to support selection for the specific application situation.

  • Guidelines: The guidelines address selecting an appropriate adaptation strategy and validating the resulting adapted models.The paper presents these as the two subsequent practical concerns.

A. How to chose the adaptation strategy

The paper recommends choosing domain-adaptation strategies according to sensor heterogeneity, available labels, existing classifiers, and the severity of dataset shift. Methods may also be combined when application conditions warrant it.

  • Sensor configuration: Different sensors create feature spaces that require heterogeneous domain-adaptation methods capable of aligning spaces with different dimensionality.
  • Available models and labels: When a source classifier already exists, classifier-adaptation and selective-sampling methods preserve that model, whereas the other families require a new classifier.
  • Available models and labels: Target labels should be acquired whenever possible because target examples provide information about the target class-conditional distribution.
  • Dataset-shift severity: Representation methods using labeled target samples can address strong nonlinear deformations, while methods without target samples require pre-aligned distributions and fail under drastic shifts.
  • Dataset-shift severity: Active learning can address larger shifts than semisupervised methods because it collects labeled target samples revealing actual target distributions.
  • Combining strategies: Invariant-feature selection, representation matching, classifier adaptation, and selective sampling can be combined according to the data and application.

B. How to validate

Validating adaptation is difficult because target labels are absent or scarce. The paper presents circular validation, which assesses target solutions through reverse adaptation using labeled source data.

  • Validation challenge: Validation is a bottleneck because adaptation assumes that no or only few labeled target samples are available.
  • Circular validation: Circular validation applies the same adaptation algorithm in reverse to assess whether an unlabeled target solution is consistent with the source domain.
  • Circular validation: Reverse adaptation uses source-domain labels to evaluate the resulting source predictions, helping determine feasibility and select an effective strategy.

V. CONCLUSIONS

The paper reviews recent remote-sensing domain-adaptation advances and organizes them into four families, pairing the taxonomy with practical examples and selection guidelines. It argues that domain adaptation is increasingly important as Earth Observation becomes multimodal and repeatedly imaged.

  • Review scope: The review groups methods into invariant-feature selection, representation matching, source-classifier adaptation, and limited target-domain sampling.
  • Future Earth Observation: Domain adaptation addresses the need to reuse ground-reference samples for new acquisitions across different areas, times, and sensors.
  • Review scope: Practical examples introduce the field and provide guidelines for selecting approaches in real application scenarios.
  • Future Earth Observation: The authors identify domain adaptation as highly important for future Earth Observation because multimodality and repeated imaging have become unavoidable.
  • Future Earth Observation: The paper presents domain adaptation and machine learning as contributing tools for trustworthy monitoring across sensor configurations or modalities.
Loading 2104.07778v1…