# A Survey of Active Learning Algorithms for Supervised Remote Sensing Image Classification

## Abstract
This survey paper provides a comprehensive overview of active learning algorithms in the context of supervised remote sensing image classification. Drawing from 100 influential research papers, it synthesizes key advancements, methodologies, and challenges, offering insights into the evolving landscape of active learning and its implications for future research. The paper highlights the integration of deep learning frameworks, uncertainty-based sampling, diversity measures, and innovative techniques such as reinforcement learning and Bayesian approaches. It emphasizes the importance of these methodologies in enhancing the efficiency and effectiveness of remote sensing applications, while also identifying critical challenges and potential future directions.

## Introduction
Active learning (AL) has become a pivotal technique in machine learning, particularly in scenarios where acquiring labeled data is costly or time-consuming. In the domain of remote sensing image classification, AL offers a promising avenue to enhance model performance while minimizing the need for extensive human annotation. This survey aims to consolidate knowledge from a vast array of studies to provide researchers and practitioners with a coherent understanding of the current landscape of active learning algorithms. The paper will focus on methodologies, results, and implications, highlighting common themes, methodologies, and advancements.

## Methodological Approaches in Active Learning

### Uncertainty-Based Sampling
Uncertainty sampling remains a cornerstone in AL strategies, where samples with the highest prediction uncertainty are prioritized for labeling. Various studies have explored different uncertainty measures, such as variance and entropy, to identify the most informative samples. For instance, Tuia et al. [12] evaluated multiple active learning algorithms and demonstrated their utility across different resolutions and spectral bands. Similarly, Růžička et al. [1] used uncertainty measures to select informative samples for change detection in remote sensing imagery. These methods are particularly effective in scenarios with high intraclass variance and complex environments.

### Diversity-Based Sampling
Diversity sampling seeks to ensure a broad coverage of the data manifold by selecting samples that are representative of different classes or regions. This approach helps in reducing redundancy and enhancing model generalization. Hekimoglu et al. [6] proposed the NORIS algorithm to ensure sample diversity in object detection, while Mackowiak et al. [5] introduced CEREALS, a cost-effective region-based active learning method for semantic segmentation. These methods are crucial in preventing overfitting and improving the robustness of models.

### Region-Level and Batch Query Strategies
Traditional AL approaches typically focus on selecting entire images or individual objects for annotation. However, region-level and batch query strategies offer more nuanced alternatives. Michael Laielli et al. [3] introduced a generalized region-level approach to avoid redundant queries and minimize context switching for labelers, significantly reducing labeling effort and improving rare object detection in cluttered scenes. Similarly, Robert Pinsler et al. [10] presented a batch active learning method that constructs diverse batches, enabling efficient learning at scale.

### Self-Consistency and Consistency Constraints
Self-consistency has been explored as a powerful source of self-supervision, particularly in semantic segmentation tasks. S. Alireza Golestaneh & Kris M. Kitani [16] utilized equivariant transformations to enforce consistency between model outputs, thereby reducing uncertainty and enhancing model performance with minimal labeled data. This approach has shown promising results, reaching nearly 96% of the performance of models trained on full datasets using only 12% of the data.

### Generative Models and Transfer Learning
The integration of generative models and transfer learning techniques represents another frontier in AL. Jia-Jie Zhu & José Bento [15] leveraged Generative Adversarial Networks (GANs) to synthesize training instances, adapting to the uncertainty principle for faster learning. Additionally, Umang Aggarwal et al. [9] employed a transfer learning-inspired approach, using pre-trained models as feature extractors and focusing on shallow classifier learning during active iterations. This method enhances robustness by favoring samples that exhibit maximum shifts towards uncertainty across consecutive models.

## Applications and Implications

### Medical Image Segmentation and Health Monitoring
Active learning has demonstrated utility in specialized domains such as medical image segmentation and health monitoring. Mélanie Gaillochet et al. [20] introduced stochastic batch querying to improve uncertainty-based sampling, yielding consistent improvements over conventional methods. Cho-Chun Chiu et al. [18] addressed the challenge of resource-constrained environments by employing a two-phased active learning approach, ensuring efficient data utilization without compromising model quality.

### Satellite Image Change Detection
In satellite image change detection, active learning facilitates interactive and efficient annotation processes. Hichem Sahbi et al. [19] developed an algorithm that models the probability of sample relevance based on representativity, diversity, and ambiguity, leading to enhanced performance in detecting changes post-natural disasters.

## Challenges and Future Directions
Despite significant advancements, several challenges persist in the field of active learning for remote sensing image classification. These include handling ambiguous and out-of-distribution samples, ensuring robustness against distribution shifts, and addressing the cold start problem. Future research should focus on developing more robust and scalable solutions that can be readily applied to a wide range of remote sensing tasks. Integrating multi-modal data and emerging technologies like federated and transfer learning could further enhance the capabilities of active learning algorithms.

## Conclusion
This survey underscores the diverse and innovative approaches to active learning in remote sensing image classification. Key themes include the integration of uncertainty sampling with diversity measures, the adoption of region-level and batch query strategies, the application of self-consistency constraints, and the incorporation of generative models and transfer learning techniques. These advancements collectively contribute to more efficient and effective annotation processes, paving the way for broader applications in remote sensing and beyond.

## References
[1] Růžička, V., D'Aronco, S., Wegner, J.D., & Schindler, K. (2022). Deep Active Learning in Remote Sensing for data efficient Change Detection.
[2] Hübotter, J., Sukhija, B., Treven, L., As, Y., & Krause, A. (2022). Transductive Active Learning: Theory and Applications.
[3] Desai, S.V., Chandra, A.L., Guo, W., Ninomiya, S., & Balasubramanian, V.N. (2022). An Adaptive Supervision Framework for Active Learning in Object Detection.
[4] Caramalau, R., Bhattarai, B., Stoyanov, D., Kim, T.-K. (2022). MoBYv2AL Self-supervised Active Learning for Image Classification.
[5] Zolfaghari Bengar, J., Gonzalez-Garcia, A., Villalonga, G., Raducanu, B., H. Aghdam, H.H., Mozerov, M., Lopez, A.M., & van de Weijer, J. (2022). Temporal Coherence for Active Learning in Videos.
[6] Hekimoglu, A., Brucker, A., Kayali, A.K., Schmidt, M., & Marcos-Ramiro, A. (2022). Active Learning for Object Detection with Non-Redundant Informative Sampling.
[7] Jose, A., Devijver, E., Amini, M.R., Jakse, N., & Poloni, R. (2022). Classification Tree-based Active Learning: A Wrapper Approach.
[8] Liu, P., Zhang, H., & Eom, K.B. (2022). Active Deep Learning for Classification of Hyperspectral Images.
[9] Bu, D., Huang, W., Suzuki, T., Cheng, J., Zhang, Q.F., Xu, Z.Q., & Wong, H.S. (2022). Provably Neural Active Learning Succeeds via Prioritizing Perplexing Samples.
[10] Mohamadi, M.A., Bae, W., & Sutherland, D.J. (2022). Making Look-Ahead Active Learning Strategies Feasible with Neural Tangent Kernels.
[11] Sinha, S., Ebrahimi, S., & Darrell, T. (2022). Variational Adversarial Active Learning.
[12] Yuan, T., Wan, F., Fu, M., Liu, J., Xu, S., Ji, X.Y., & Ye, Q.X. (2022). Multiple instance active learning for object detection.
[13] Rakesh, V., & Jain, S. (2022). Efficacy of Bayesian Neural Networks in Active Learning.
[14] Miller, K., Mauro, J., Setiadi, J., Baca, X., Shi, Z., Calder, J., & Bertozzi, A.L. (2022). Graph-based Active Learning for Semi-supervised Classification of SAR Data.
[15] Rangnekar, A., Kanan, C., & Hoffman, M. (2022). Semantic Segmentation with Active Semi-Supervised Learning.
[16] Bouneffouf, D. (2022). Exponentiated Gradient Exploration for Active Learning.
[17] Yu, W.P., Zhu, S.J., Yang, T.J.N., & Chen, C. (2022). Consistency-based Active Learning for Object Detection.
[18] Bickford Smith, F., Foster, A., & Rainforth, T. (2022). Making Better
[19] Takezoe, R., Liu, X., Mao, S., Chen, M. T., Feng, Z., Zhang, S., & Wang, X. (2020). Deep Active Learning for Computer Vision: Past and Future.
[20] Tuia, D., Volpi, M., Copa, L., Kanevski, M., & Munoz-Mari, J. (2020). A survey of active learning algorithms for supervised remote sensing image classification.