Little learning in remote sensing: trends, gaps, and future trends

Links table

1. Abstract
2. backgrounds
3. Remote sensing data type
4. Standard measure
5. Rating measures for a few shortcomings
6. The last few learning techniques in remote sensing
7. Detection of snapshots and retailers of remote sensing
8. Discussions
9. Numerous experimenting to classify a few aircraft on the drone -based data collection
10. AI (Xai) can be explained in remote sensing
11. Future conclusions and trends
12. Thanks, appreciation, ads and references

8 discussions

In this section, we aim to shed light on interesting notes, common trends and potential research gaps based on the in -depth analysis of current few classification techniques through the areas of sensing data for the three dimensions. The visions discussed in this section can be a guide for all current and future researchers in this field.

• Most of the methods described in literature use various features of extracting features, as CNN models are often the spine, as we have already talked. The few few learning models are still common for classification tasks in all three fields. These models are able to adapt quickly to new seasons with a few training examples, making them suitable for applications in the real world. However, the methods based on the graph have become more popular to classify SAR, and they were only used to classify VHR photos. The methods based on the graph are useful because they are able to capture spatial relationships between organisms, which is necessary to classify SAR and VHR images. Recently, learning -based methods have emerged on the vision transformer and learning as alternatives to classifying high -spectrum images. These methods have shown a promise to achieve high accuracy with the minimum training data, making them attractive to the applications in which the name is limited.

• The business evaluation discussed in the classification of high -spectrum images is used in general three common use: overall resolution (OA), average accuracy (AA), and Kappa (κ). These scales are used frequently to evaluate the performance of proposed algorithms. In contrast, for VHR and SAR images, classification accuracy (OA) is often used as a basic evaluation scale, although there are few exceptions. Moreover, in most of the evaluation strategies adopted by researchers, proposed algorithms are operated several times in addition to newer techniques (SOTA), and average medium accuracy and standard deviation are reported. This approach provides a more reliable and powerful appreciation for the performance of the classification, taking into account any possible differences in the results obtained through multiple operations.

• Unlike the super -spectral classification, it is currently noted that there are currently a few rating methods of vision or non -vision in Vision proposed for SAR and VHR images. This can be attributed to the challenges associated with the acquisition of sufficient data sets to implement effective and accurate structures based on SAR. Likewise, for VHR images, although there are models that use the category based on VIT, they are unexpected curricula as the vanilla -based model suggested by Zhang et al. [3]. Consequently, there are great opportunities for researchers to explore the capabilities of the few methods that depend on the few shot to face the challenges associated with the classification of remote sensing data.

• It does not seem that the current situation to search for a few classification approaches in the field of remote sensing includes a lot of work on pictures of drones or low aircraft images, as much as the current knowledge suggests. This may be due to the unique nature of such images, which have been referred to in previous studies such as [95]. The differences in the sizes of organisms and views, as well as the limited mathematical resources available for drones -based operations, may be contributing factors in the scarcity of research in this field. In addition, the relatively smaller volume of drone -based data groups may be challenges to few learning methods, which often require a large enough data set to learn representations of meaningful feature. However, with increased availability of data -based data, there may be opportunities to develop a few new classification methods of shots that can effectively benefit from these data.

• Moreover, although learning is a little bit of learning widely in the context of the classification subject to supervision, there is also a possibility to explore its application in other remote sensing tasks such as not subject to supervision or semi -noticeable, detection of organisms, and semantic fragmentation. Learning can provide a little learning an effective way to take advantage of the limited data called these tasks, which can lead to more accurate and effective algorithms for remote sensors. In general, although great progress in the application of learning a little shots on remote sensing data is still many research gaps and opportunities for further investigation. Exploring new learning methods, as well as extending the current methods of new applications and fields, can lead to more accurate and effective algorithms of remote sensing tasks.

• The use of Xai methodologies in conjunction with the few learning models for remote sensing applications can significantly enhance the interpretation of these models significantly, and thus increases the ability to apply them in areas that are sensitive to potential risks. However, despite the great promise that Xai keeps learning a little in remote sensing, the current research group in this field is still relatively emerging, and other endeavors are necessary to achieve its potential benefits completely.

8.1 Account considerations in learning a few

Learning a little, as a place in the broader field of automatic learning, calls for unique mathematical requirements. These requirements become particularly relevant when applications have actual time restrictions. One of the most actual applications is to monitor disaster with drones. The immediate comments in such scenarios can significantly affect the results, stressing the importance of the treatment time.

Deep learning, which constitutes the basis for many learning techniques, requires high mathematical resources. Techniques like

CNNS is famous for its arithmetic intensity during both training and reasoning. This mathematical cost can sometimes be the bottleneck, especially when fast responses are necessary. However, the sophisticated scene of learning a little witnessed in the emergence of strategies aimed at alleviating these mathematical challenges:

• Paining learning, it is clear from approaches such as maml [125]It provides an innovative solution. By improving the model parameters to allow rapid adaptation to new tasks, these methods significantly reduce the general account expenditures. This ensures that models can be efficiently adjusted, even when facing new data collections.

• Wang and others [70] The proposal to use a lightweight model structure is highlighted as well as distillation techniques as a viable and viable strategy. By reducing unnecessary repetition and parameters, these models are simplified to be more effective in terms of mathematical aspects without prejudice to their predictive authority.

• Graphsage methodologies, such as Graphsage [50]And more extensions in GNN approaches [80]Providing alternatives to the traditional CNN network. These methods, in specific data collection contexts, showed a decrease in arithmetic complexity, making them attractive options.

Despite these developments, it should be noted that a large part of the few learning curriculum has not been explicitly designed to improve treatment time. When you realize this gap, future research can tend to formulate specially designed structures for drone applications in actual time. Several ways to enhance mathematical efficiency can be pursued. These include the adoption of model pressure techniques, such as pruning and supplementary [126]Take advantage of the effective research methods of nervous architecture [127]And explore the joint software design strategies [128] To adjust the models for certain mathematical platforms. In all these endeavors, the comprehensive goal remains constant: achieving rapid conclusion times without sacrificing the accuracy of the model.