SYSTEMS AND METHODS FOR GRAPH-BASED ACTIVE LEARNING USING SEMI- SUPERVISED CLASSIFICATION OF SAR DATA GOVERNMENT LICENSE RIGHTS [0001] This invention was made with government support under DMS-1944925, DMS- 1952339 and DMS-2027277 awarded by the National Science Foundation. The government has certain rights in the invention. [0002] This invention was made with government support under Grant No. HM04762110003 awarded by the Department of Defense/National Geospatial Intelligence Agency. The government has certain rights in the invention. CROSS-REFERENCE TO RELATED APPLICATIONS [0003] The current application claims the benefit of and priority under 35 U.S.C. § 119(e) to U.S. Provisional Patent Application No.63/362,385 entitled “Systems and Methods for Graph-Based Active Learning for Semi-Supervised Classification of SAR Data” filed April 1, 2022, the disclosure of which is hereby incorporated by reference in its entirety for all purposes. FIELD OF THE INVENTION [0004] The present invention generally relates to automatic target recognition using supervised machine learning algorithms. BACKGROUND OF THE INVENTION [0005] As data becomes more conveniently collected and stored, large amounts of unlabeled data give data scientists more data than they are capable of analyzing. Semi- supervised learning (SSL) is an approach to machine learning that involves a small amount of labeled data with a large amount of unlabeled data to achieve an accurate classification with significantly fewer training points. [0006] Active learning is a subset of SSL which interactively queries a human agent to label a selected number of unlabeled datapoints with the desired outputs Since manual learning is becoming an appealing option to process these unlabeled data to improve underlying SSL classifier’s performance. [0007] Synthetic Aperture Radar (SAR) is a form of radar that utilizes the motions of a radar antenna over a distance from the target region to create two-dimensional images or three-dimensional reconstructions of objects. SAR is able to provide finer spatial resolution than conventional stationary beam-scanning radars. SUMMARY OF INVENTION [0008] Systems and methods for graph-based active learning using semi- supervised classification of synthetic aperture radar (SAR) data in accordance with embodiments of the invention are illustrated. One embodiment includes an automatic target recognition system using digital images and active learning, including a display device, and a computing device, wherein the computing device includes at least one processor, a memory, and at least one non-transitory computer-readable media comprising program instructions that are executable by the at least one processor such that the computing device is configured to perform automatic target recognition by receiving synthetic aperture radar (SAR) images, transforming the SAR images to a lower dimensional space to extract feature vectors, constructing a similarity graph based on the feature vectors, training a neural network model using the similarity graph, and performing a graph-based active learning process with the trained neural network model to classify and identify targets of interest in the SAR images. [0009] In another embodiment, the graph-based active learning process includes applying Laplace learning to infer a plurality of labels of a plurality of sets of unlabeled data based on the similarity graph and a small plurality of labelled data, adding the plurality of inferred labels to an underlying semi-supervised classifier, identifying a query point consisting of unlabeled data, determining whether the query point should be labeled, displaying the query point on the display device when it is determined that the query point should be labeled, capturing input indicating that the query point represents features of the SAR image adding the captured input to the underlying semi-supervised classifier and iteratively repeating the adding of label and capturing of input to the underlying semi- supervised classifier until no unlabeled points remain. [0010] In a further embodiment, the graph-based active learning process includes computing a core set of initial labels based on similarity graph, adding the core set of initial labels to an underlying semi-supervised classifier, identifying a set of query points consisting of unlabeled data, determining whether the set of query points should be labeled, displaying the set of query points on the display device when it is determined that the set of query points should be labeled, capturing inputs in parallel indicating that at least one point from the set of query points represent features of the SAR image, adding the captured input to the underlying semi-supervised classifier, and iteratively repeating the adding of labels and capturing of input to the underlying semi-supervised classifier until no unlabeled points remain. [0011] In still another embodiment, determining whether the query point should be labeled includes using an acquisition function. [0012] In a still further embodiment, the acquisition function includes an uncertainty acquisition function. [0013] In yet another embodiment, determining whether the set of query points should be labeled includes using a LocalMax function. [0014] In a yet further embodiment, using the LocalMax function includes selecting a query set of points that satisfies a local maximum condition on the similarity graph. [0015] In another additional embodiment, the neural network model is a contrastive learning model. [0016] In a further additional embodiment, the trained neural network is based on a PyTorch convolutional neural network. [0017] One embodiment includes a method for automatic target recognition using digital images and active learning, where the method includes: receiving synthetic aperture radar (SAR) images, transforming the SAR images to a lower dimensional space to extract feature vectors, constructing a similarity graph based on the feature vectors, training a neural network model using the similarity graph. And performing a graph-based active learning process with the trained model to classify and identify targets of interest in the SAR images. [0018] Additional embodiments and features are set forth in part in the description that follows, and in part will become apparent to those skilled in the art upon examination of the specification or may be learned by the practice of the invention. A further understanding of the nature and advantages of the present invention may be realized by reference to the remaining portions of the specification and the drawings, which forms a part of this disclosure. BRIEF DESCRIPTION OF THE DRAWINGS [0019] The description and claims will be more fully understood with reference to the following figures and data graphs, which are presented as exemplary embodiments of the invention and should not be construed as a complete recitation of the scope of the invention. [0020] Fig.1 illustrates a process for training models under semi-supervised learning (SSL) in accordance with embodiments of the invention. [0021] Fig. 2 illustrates an example contrastive learning process in accordance with embodiments of the invention. [0022] Fig. 3 illustrates an example transfer learning process in accordance with embodiments of the invention. [0023] Fig.4 illustrates an example process of sequential active learning in accordance with embodiments of the invention. [0024] Fig.5 illustrates an example process of batch active learning in accordance with embodiments of the invention. [0025] Fig.6 illustrates an algorithm used to compute the core-set in accordance with an embodiment of the invention. [0026] Fig.7 illustrates an algorithm for LocalMax Batch Active Learning in accordance with an embodiment of the invention. [0027] Fig. 8 illustrates a high-level block diagram of a system that the processes described above can be implemented on in some embodiments of the invention [0028] Fig.9 illustrates a comparison between visualizations of embeddings extracted using contrastive learning. [0029] Figs.10A-B illustrates the accuracy of active learning using contrastive learning embeddings. [0030] Fig.11 illustrates the accuracy of active learning with Laplace semi-supervised learning on the VAE embeddings with the pretrained weights. [0031] Fig.12 illustrates a direct comparison between the graph based active learning performance with the contrastive learning embeddings against the VAE embeddings. [0032] Fig.13 demonstrates the time consumption and accuracy comparison among different active learning methods. [0033] Figs.14A-B illustrates an accuracy comparison with respect to the number of labeled points in two data sets for five different active learning methods. [0034] Fig.15 illustrates an accuracy comparison with respect to the number of labeled points for each embedding and dataset. DETAILED DESCRIPTION OF THE DRAWINGS [0035] Automatic target recognition (ATR) refers to the ability for a system or device to recognize targets based on data obtained from sensors. Synthetic Aperture Radars (SARs) are a type of imaging radar that is mounted on a moving platform, and repeatedly transmits and receives radio signals to simulate a large radar dish and achieve high resolution images. SAR imaging has a high level of reconstruction accuracy, which allows it to have a wide range of applications in identification of specific targets, particularly for military applications. However, difficulties arise with classifying images collected by SARs. Even though the images captured are of higher resolution, objects within the captured images may be difficult for a human to classify, and may often require experts in the domain to properly identify the objects and classify the images. This in turn, encourages the use of machine learning methods to improve speed and accuracy of object identification for SAR images. [0036] Current methods for image classification rely mainly on supervised learning, which requires a significant amount of labeled data to train machine learning models such that the model is able to classify images correctly. However, there are few labeled data points in SAR images. Therefore, supervised learning is less suitable for SAR image classification. [0037] Systems and methods described herein attempt to remedy the above issue by introducing an image classification method for automatic target recognition in SAR images that uses semi-supervised learning (SSL) in combination with unsupervised learning. SSL can leverage the fact that there are limited number of labeled points in SAR image data sets to train machine learning models for image classification. In many embodiments, the system extracts feature vectors from input data that have a limited number of labels, and constructs a similarity graph based on the extracted feature vectors. As part of the SSL process, the system may train machine learning models using the similarity graph. In numerous embodiments, the trained models are deployed in an iterative active learning process to identify unlabeled points. Oracles, who are experts in a domain depicted by the SAR images, may determine whether the unlabeled points should be labeled to assist in the learning of models. In several embodiments, an input data set is updated with the newly labeled data points and returned back to the model to identify further unlabeled points. In some embodiments, the active learning process repeats until no unlabeled points remain, which indicates that the models are optimized for SAR image classification. [0038] Original SAR images can be of various sizes, and therefore in many embodiments, the magnitude and phase images are normalized to a similar size, including by center-cropping to 88x88 pixels in certain embodiments. In several embodiments, magnitude images are then clipped to the range [0,1], and each magnitude-phase image pair is transformed into a 3-channel image. This transformation may be necessary for the loss function in certain variational autoencoders. It should be noted that the SAR images used in this process and system can be replaced with hyperspectral images, mass spectroscopy images, or LiDAR images in various embodiments of the invention, and the system can still produce similar results. [0039] A process for training models under semi-supervised learning (SSL) in accordance with embodiments of the invention is illustrated in Fig 1 Process may be implemented on an SAR image classification system, such as those that will be described further below. Process 100 receives (110) input data consisting of images collected by SARs. In several embodiments, input data contains a limited number of labels. Process 100 can extract (120) feature vectors from input data. Feature vectors may be generated by encoding the input data to a lower dimensional space. Process 100 constructs (130) similarity graph based on the feature vectors, and trains (140) prediction model using the constructed similarity graph. Techniques for extracting feature vectors, constructing similarity graphs, and training prediction models using a constructed similar graph that may be utilized in accordance with embodiments of the invention will be described in greater detail below. In many embodiments, any of these techniques may be incorporated into processes such as process 100 described above with respect to Fig.1. Feature Extraction [0040] In many embodiments, feature extraction is performed by a contrastive learning model. A contrastive learning model extracts features by minimizing the distance of like- images and maximizing the separation of “unlike” images in the encoding space. The images that are designated as similar samples can be referred to as positive pairs whereas the images that are designated as dissimilar can be referred to as negatives. In many embodiments, a series of augmentations are performed on input data such that each image will have two different resulting images after the augmentations. An example contrastive learning process in accordance with embodiments of the invention is illustrated in Fig.2. In several embodiments, augmenting an image xi can produce two augmented images denoted by xi1 and xi2 respectively. Since these two augmented samples x _i1 and x _i2 , were both derived from the same image x _i , they are generally considered as positive pairs and all the other augmented images in the batch may be considered as negative pairs. [0041] In many embodiments, cosine similarity function is used to measure the similarity between the features extracted from two images, u and v, where the cosine similarity function is: _{^ ^} ^{For a given positive pair, x} _i1 ^{, x} _i2 ^{, the} ₁ where τ is a temperatu the similarity function. The loss, summed over an entire batch, is called NT-Xent, which stands for normalized temperature-scaled cross entropy loss. The loss for a batch of n samples is: [0042] In several embodiments, each on e o t e augmented mages w be passed through the first phase of the neural network which is an encoder. Encoders are convolutional layers that may perform feature extraction by transforming an augmented sample x _ij into a feature embedding denoted as h _ij . A projection head, which is typically a multi-layer perceptron with one or two hidden layers may be used to transform a feature embedding hij to a new space where the outputs are denoted as zij . The NT-Xent loss function may be evaluated on the projected features z _ij and the encoder and projection head networks are simultaneously updated towards minimizing the loss. [0043] Augmentation should be done in moderation as too harsh an augmentation and the network may learn from noise and undesirable artefacts of the transformation. If there are too few augmentations, the network may not generalize well. Contrastive learning frameworks for image classification (e.g. ImageNet, Cifar10) typically use color jitter, random cropping, random horizontal or vertical filps, and random blur for augmentations. But SAR images are generally taken from airplanes where the scanning vehicles cast a shadow where no radar signals were received. This shadow is always behind the vehicles, and in selected embodiments, the encoder model is configured to learn this feature. Therefore, augmentations for the dataset that would rotate or flip the images vertically were not performed, as this would destroy the shadow effect always being behind the target. [0044] Additionally, SAR images typically are not RGB images and instead have magnitude and phase information, therefore color jitter augmentation may be omitted. Finally, a random center crop augmentation can be used where an integer 40 ≤ k ≤ 88 is randomly selected, and the image is then cropped around the center to produce a k × k image that would be resized to 32 × 32. These adjustments can build scale and zoom invariance. The downsizing of the image to 32 × 32 can mitigate memory usage issues and the dimension being a power of 2 allows for several max-pooling CNN layers, granting the encoder greater capacity for abstractification and less susceptibility to pixel-wise minutia. The other two standard data augmentations, either of which may be applied to the input data, are a random horizontal flip, which flips the image horizontally 50% of the time and a random Gaussian blur transformation that uses a 7 × 7 kernel and a random sigma value randomly selected between 0.1 and 2.0 for each augmentation. [0045] In many embodiments, the encoder is a standard ResNet32 architecture and the projection head is a 3-layer perceptron. In order to have large batch sizes under memory constraints and prevent overfitting, in some embodiments, a ResNet18 model can be used. The ResNet18 can be trained for 500 epochs and 1000 epochs with a learning rate of 0.05, batch size of 512, and a SimCLR temperature of 0.5. [0046] In numerous embodiments, feature extraction can be performed using transfer learning based on pre-trained PyTorch convolutional neural networks (CNN). An example transfer learning process in accordance with embodiments of the invention is illustrated in Fig.3. In many embodiments, the encoder model used in transfer learning is made up of many convolutional layers, with a final convolutional layer near the end of the neural network designated as the feature layer since it captures complex features of the dataset. The encoder portion of the neural network is primarily composed of 2D convolutions that compress the images to a low dimensional space. Transfer learning involves reusing of a neural network trained on a similar dataset and task for a new task. The parameters in the convolutional layers of the pretrained CNN are transferred to the new CNN. In fine-tuned transfer learning all the transferred parameters are trained for a few iterations on the new dataset. In several embodiments, the new CNN is trained by first adding new fully connected layers at the end of the neural network and performing supervised learning with the new dataset. When training is complete, only the convolutional layers are used for the embedding process. Similarity graph [0047] With the features extracted, image data may be viewed as vectors embedded in a much lower dimensional space, R ^d , than the raw image data. Quantitative comparisons can be performed on and between the image feature vectors extracted using the contrastive learning model to construct a similarity graph for the dataset. In many embodiments of the invention, the cosine similarity function may be used to construct a k nearest neighbors graph from the embedded images. [0048] In several embodiments, a fast nearest neighbors approximation algorithm is performed to boost computation. Once the k nearest neighbors for each sample are selected, the edge weights in the resulting graph can be determined. A self-tuning similarity graph can be constructed with edge weights given by: _{where xi, xj represent the normalized features for the i} ^th _{and j} ^th _{images respectively, and} ^d _k ^(x _i ^{) represents the distance between x th} i ^{and its k} _{nearest neighbor.} [0049] In some embodiments, similarity graph G = (X ,W) for features extracted using transfer learning is constructed where X is a set of d-dimensional feature vectors X = {x ₁ , x ₂ ,..., x _N } ⊂ R ^d and W∈R ^N×N is the weighted adjacency matrix. W _ij denotes the weight on the edge between vertices xi, xj, i ≠j which measures the similarity between feature vectors xi and xj. The weight matrix can be defined by: where d(x _i , x _j ) is the distance between feature vectors x _i and x _j . f is the kernel function and σi and σj are normalization constants corresponding to xi, xj, respectively. Distance function d in many embodiments is the angular distance: The normalization constant σi associated to node i is cho the K ^th nearest neighbor of i , where xiK is the K th nearest neighbor to xi . The kernel function f is the Gaussian (exponential) kernel f (x) = exp(−x). Graph-based SSL [0050] In many embodiments, semi-supervised learning (SSL) based on constructed similarity graphs uses a set of labeled data L and a weighted adjacency matrix W corresponding to the graph. Let {x1, ..., xn} := X ⊂ R ^d represent the feature vectors of the images in d dimension and let L ⊂ {1, 2, ..., n} represent the set of indices for feature vectors that have an associated, known label. U = {1, 2, ..., n}−L represents the set of indices of unlabeled images. For a dataset with K classes, yj ∈ {1, 2, ..., K} represent the label for image j so that the labels correspond to their one-hot encoding as eyj ∈ R ^K . Let G(X , W) be a graph where X represents the vertices and W represent the edge weights between the images. In some embodiments, Laplace learning solves for the label function f : X → R ^K such that: where d _i = ^∑ _{j ¹} W _ij This solution can generate a graph harmonic function which propagates the labels across the dataset. [0051] In some embodiments, the SSL scheme used is Laplace learning, where the predicted label of an unlabeled node x _i ∈ X\X _L is denoted by . With information on the ground truth labels for the subset X _L ⊂ X, the goal for the SSL in several embodiments of the invention is to predict the labels of the unlabeled nodes x _i ∈ X\X _L . In many embodiments, SSL Laplace learning involves identifying whether the j ^th entry of vector u(x _i ) belongs to class j. In several embodiments, identification requires the model to obtai n a classifier by identifying ^^^ with its matrix representation, and solving for where < ., .>F is e o e us e po uc o a ces. s e u o a e graph Laplacian matrix, where D is the diagonal matrix with diagonal entries ^^ _^,^ ൌ ∑ _k∈Z W _jk for 1 ≤ j ≤ N. Sequential Active Learning [0052] In numerous embodiments of the invention, the model trained under graph-based SSL undergoes further learning to better leverage the limited amount of labeled data. In this phase, the model goes through an iterative sequential active learning process where acquisition functions may be applied to the predictions from the trained model to determine which unlabeled sample, i ∈ U should be labeled by an oracle, referring to a human expert in the domain depicted by the SAR images. An example process of sequential active learning in accordance with embodiments of the invention is illustrated in Fig.4. [0053] Process 400 identifies (410) query points in a data set that are predicted by the SSL-trained model. Process 400 determines (420) whether the queried points should be labeled. In many embodiments, this determination is performed using acquisition functions including the uncertainty acquisition function, among others. The uncertainty acquisition function can query the sample that lies closest to the decision boundary, which represents the border between expected classes in the data. In many embodiments, the uncertainty acquisition function utilizes prediction models to process unlabeled nodes, and outputs a vector of length K such that the i ^th component corresponds to the likelihood that the given sample belongs to the i ^th class. In practice, a threshold may be put on the predictions so that the model can directly output class predictions. By using the soft prediction vector, a notion of confidence and uncertainty in how the model predicts certain classes can be established. Heuristically, the uncertainty acquisition function will be able to target regions of the dataset that are low confidence and allows for refinement. For a model prediction v∗(x _i ) ∈ R ^K for the sample i ∈ U, the margin may be defined as the difference between the first and second highest predictions: where y∗(i) is the la ges pe c o aue u . [0054] In some embodiments, variance optimality (VOPT) may be used as the acquisition function. In several embodiments, the Model Change function can be used as an acquisition function for active learning. The goal of active learning is to select the next points that would be most helpful for the model to obtain labels for. [0055] Process 400 classifies (430) the queried points that should be labeled. Points may be queried one at a time. Step 430 may involve input from an expert oracle to accurately classify the points chosen by the model. Process 400 updates (440) the labeled data set with newly labeled points from the oracle. Process 400 determines (450) if there are any unlabeled data remaining, and if there are unlabeled data, the updated data set is returned (460) to the model for additional queries. Process 400 iteratively repeats until there are no unlabeled points remaining. [0056] In many embodiments, the updated data set can be used to retrain or fine-tune the prediction model. Typically, active learning is preferred when either individual samples require significant time and energy to label or there are far too many samples to effectively label. SAR data falls under the first category, as the noisy images are nearly incomprehensible to an untrained human. Batch Active Learning [0057] When the data set is sufficiently large, the active learning phase may use a batch active learning method to be more efficient In batch active learning the model can select multiple points to make up a query set for multiple oracles and/or teams to classify in parallel. This reduces the total labeling time by an order of magnitude as compared to the sequential case. An example process of batch active learning in accordance with embodiments of the invention is illustrated in Fig.5. Process 500 computes (510) a core- set based on a constructed similarity graph. [0058] In many embodiments, an algorithm iteratively selects nodes in X in graph G = (X,W) to construct a core-set such that all points are a distance at least r from each other but no more than distance R from another point. At each step of the core-set selection process, assuming the currently selected node set is Y (i.e. current labeled set), the algorithm creates an annular set C and a seen set S: where ^ _{11^ and d (x, y ) is the dist} G _{ance from x to y computed using Dijkstra’s algorithm. The annular set} is the set of points that may be selected from at each stage in the algorithm and the seen set is a set of points that the algorithm can no longer select from. The algorithm may randomly select x ∈ C and updates Y, S, C. Repeating this process can give a uniform covering of the data. Process 500 identifies (520) a set of query points from input data. In many embodiments, step 520 relies on a novel batch active learning approach named LocalMax as the acquisition function. LocalMax may select a query set of multiple nodes that satisfies the local maximum condition on the constructed similarity graph. The local maximum condition is defined by: Assuming a batch size B , at iteration k LocalMax selects the query set Q _k as the top-B local maximums in the data set. [0059] Process 500 classifies (530) the set of query points. In many embodiments, the set of query points may be classified in parallel by teams of oracles to ensure efficiency. Classifying with teams of oracles can allow the active learning to avoid the largest bottleneck in the process. Process 500 updates (540) labeled data set with the classified query points. Process 500 determines (550) if there are any unlabeled data remaining, and if there are unlabeled data, the updated data set is returned (560) to the model for additional queries. Process 500 iteratively repeats until there is no unlabeled points remaining. [0060] Fig. 6 illustrates an algorithm that may be used to compute the core-set in accordance with an embodiment of the invention. Fig. 7 illustrates the algorithm for LocalMax Batch Active Learning in accordance with an embodiment of the invention. By building an acquisition function on sequential active learning allows LocalMax to benefit from many useful properties including simplicity and efficiency. Controlling for local maxes enforces a minimum pairwise distance between points in the query set, which counteracts the redundancy seen in naively optimizing sequential acquisition functions. LocalMax also maintains good computational complexity. [0061] Fig. 8 illustrates a high-level block diagram of a system that the processes described above can be implemented on in some embodiments of the invention. Computing device 800 includes an input/output interface (810) that can receive the SAR image input, a memory (830) including a model application (832) and a data storage (834) to store the data sets. Device 800 further includes a processor (820) that may run and train the model (832). Input/output interface (810) can be, for example, an interface to local storage (e.g., hard disk, SSD, etc.), network-attached storage, or remote storage (e.g., cloud). Additionally, the system also includes a display device (840) to display the unlabeled points to the oracle for further labeling. Results [0062] Fig.9 illustrates a comparison between visualizations of embeddings extracted using contrastive learning. The bottom row shows embedding extracted with contrastive learning models trained for 500 epochs Colors correspond to samples of the same class The SAR image embeddings exhibit some structure, but this is largely obscured; the classes are largely intermixed and non-connected. The VAE embeddings exhibit greater class cohesion with more connected strands, but intermixing is still evident. With contrastive learning, individual classes are far more connected and better separated from others. Visualizations from contrastive learning demonstrate more cohesion within the labeled clusters. There is very little mixing and the labeled clusters are very well connected with their respective classes. [0063] Figs.10A-B illustrate the accuracy of active learning using contrastive learning embeddings. Using 300 labels, the 500 epoch embeddings achieved an average accuracy of 98.3% and the 1000 epoch embeddings achieved an average accuracy of 99.2%. Fig.11 illustrates the accuracy of active learning with Laplace semi-supervised learning on the VAE embeddings with the pretrained weights, and with 300 labels, the highest accuracy achieved is 94.2%, which is lower than its contrastive learning counterpart. Fig. 12 illustrates a direct comparison between the graph based active learning performance with the contrastive learning embeddings against the VAE embeddings. The contrastive learning embeddings are trained to 1000 epochs and averaged over 21 distinctly trained models, with the vertical bands corresponding to 1 standard deviation in accuracy. [0064] Fig.13 demonstrates the time consumption and accuracy comparison among different active learning methods. This experiment uses a CNNVAE embedding for MSTAR and zero-shot transfer learning for OpenSARShip and FUSAR-Ship data. All batch active learning methods have comparable efficiency and are 9 to 15 times faster than the sequential case. The local max method always achieved higher accuracy than other batch active learning methods and is comparable to the accuracy of sequential active learning. [0065] Figs. 14A-B illustrate an accuracy comparison with respect to the number of labeled points in two data sets for five different active learning methods. In each figure, the LocalMax method and the sequential active learning are almost identical and are the best performing methods with Local-Max being the more efficient, proportional to the batch size Fig 15 illustrates an accuracy comparison with respect to the number of labeled points for each embedding and dataset. Each panel shows four curves generated by LocalMax using different active learning acquisition functions, uncertainty, VOpt, model-change and model-change VOpt in comparison with the state-of-the-art, CNN- based method. The uncertainty acquisition function has the best performance among all the acquisition functions tested. [0066] Although specific methods of graph-based active learning using semi- supervised classification of SAR data are discussed above, many different methods of graph-based active learning using semi-supervised classification of SAR data can be implemented in accordance with many different embodiments of the invention. It is therefore to be understood that the present invention may be practiced in ways other than specifically described, without departing from the scope and spirit of the present invention. Thus, embodiments of the present invention should be considered in all respects as illustrative and not restrictive. Accordingly, the scope of the invention should be determined not by the embodiments illustrated, but by the appended claims and their equivalents.