Method and system for automatic detection of fairy circles

Abstract

A method for finding sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource, the method including receiving (900) target images (401) associated with an area of interest, generating (902) pseudo-labels (506) with a foundation model (400) by prompting the foundation model (400) with input images (502), which are different from the target images (401), training (904) a semantic segmentation model (500) based on the pseudo-labels (506), generating (906) SSD predictions (510) with the semantic segmentation model (500), from the target images (401), and producing (908) an image of the area of interest, based on the SSD predictions (510), to identify a location of the resource.

Description

METHOD AND SYSTEM FOR AUTOMATIC DETECTION OF FAIRY CIRCLESBACKGROUND OF THE INVENTIONTECHNICAL FIELD

[0001] Embodiments of the subject matter disclosed herein generally relate to a system and method for processing plural images for extracting information that might indicate the presence of a formation (e.g., geological feature) located in the subsurface, and more particularly, to one or more machine learning-based processes for determining the location of a sub-circular surface depression (SSD), also known as surface depression, circular depression, subsurface depression, semicircular depression, ovoid or fairy circle, which might be indicative of the geologic feature, for example, a gas reservoir in the subsurface.DISCUSSION OF THE BACKGROUND

[0002] SSDs are surface features that have been observed in various parts of the world. They are linked to several phenomena including, but not limited to, bacterial activity, permafrost, termite activity as well as self-organizing grasses and vegetation competing for limited water and nutrients [1 , 2, 3, 4], Notably, SSDs have also been identified, on several occasions, as a result of natural hydrogen migrating from subsurface systems to the surface [2, 5]. Therefore, SSDs are considered a coarse indication of the presence of native hydrogen in the subsurface and can contribute to a play-based hydrogen exploration de-risking strategy.

[0003] As the hydrogen fuel holds the potential to offer significant environmental and human health benefits, primarily by addressing issues related toair quality, reducing greenhouse gas emissions, and promoting sustainable energy practices, finding natural hydrogen reservoirs is highly desired. One advantage of the hydrogen fuel is the improved air quality. Hydrogen could be used with fuel cell systems to produce electrical energy, for use or storage. The combustion of hydrogen does not produce harmful pollutants such as particulate matter, or carbon monoxide (CO) or carbon dioxide (CO2). This is particularly relevant for urban areas experiencing air pollution, as the use of hydrogen fuel in transportation and industries can lead to cleaner air, reducing respiratory illnesses, and other health problems associated with poor air quality.

[0004] Future hydrogen exploration might resemble the present hydrocarbon exploration and development, which uses waves (e.g., seismic waves or electromagnetic waves) to explore the structure of underground formations on land and / or at sea (i.e., formations under the seafloor). Based on these waves, it is possible to map the structure and character of geological formations underground using, for example, reflection seismology.

[0005] However, a seismic surveying campaign employed for collecting the seismic data is expensive and time consuming. With the availability of powerful computers, applications of deep learning have brought about a revolution to various fields including geology. Studies have shown the potential of using deep neural networks (DNNs) for detecting geological bodies on seismic images, such as faults and salt bodies, by treating this task as one of image segmentation.

[0006] All these methods require the acquisition of seismic data, which is an expensive and tedious process. In some instances, determining where to perform seismic acquisition for identifying the location of a resource underground ischallenging. To increase the confidence level where the hydrogen reservoirs or hydrogen accumulations are located and / or where the seismic acquisition could be performed, SSD detection may be performed as part of an exploration program.

[0007] However, finding and mapping SSDs over vast expanses of terrain can be time and resource consuming. For this reason, machine learning based approaches have been adopted to automate the process and make it faster and more scalable. Lundine and Trembanis [6, 7] used digital elevation model (DEM) data to train Faster-RCNN, Mask-RCNN and YoloV5 models to detect the Carolina Bays in Delaware, North Carolina. Leong et al. [8] similarly trained a YOLOX object detector to map the SSDs in Southwest and Southeast Australia using a combination of Sentinel 2 bands and DEM. While possibly not related to natural hydrogen, SSDs located in Namibia have also been mapped and studied. Zhu et al. [9] used high resolution images of the Namibia circles captured by an unmanned aerial vehicle to train an image classification neural network. Noy et al.

[0010] conducted a comprehensive data collection and analysis study which included high resolution WorldView 2 satellite images with derived indices and texture characteristics to train a UNet model for SSD segmentation. They also used geometric descriptors, terrain attributes, and performed field measurements and analyses to further study the spectral and spatial properties of the circles.

[0008] The current approach when dealing with SSDs is to rely on traditional supervised learning approaches to address this problem. The workflow associated with such an approach, which is schematically illustrated in FIG. 1 , includes a step 100 of selecting an area of interest (AOI) in a given geographical area of the globe, and then, the necessary imagery (of the AOI) is acquired and processed in step 102.Several SSDs are visually identified and annotated by an experienced geologist in step 104. These SSDs are selected from the AOI. A known semantic segmentation model is then trained in step 106 using these annotations. In step 108, the model may be further refined and in step 110, the optimized model is applied to the recorded data 102 for determining SSDs in the AOI.

[0009] This approach effectively uses supervised machine learning to map SSDs. All these approaches focus only on one location (the AOI) with a single type of SSDs. However, the SSDs appear in a variety of environments and landscapes and adopt different morphological and spectral aspects. This domain shift makes a machine learning model trained in one location deficient for another location.Because the current state of the art relies on manual dataset annotation (see step 104 in FIG. 1), it needs to be repeated for every new AOI, which is both time and resource consuming.

[0010] Thus, there is a need for a new workflow for detecting SSDs no matter the region in which they developed, and for avoiding extensive training, to overcome the above noted problems.SUMMARY OF THE INVENTION

[0011] According to an embodiment, there is a method for finding sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource. The method includes receiving target images associated with an area of interest, generating pseudo-labels with a foundation model by prompting the foundation model with input images, which are different from the target images, training a semantic segmentation model based on the pseudo-labels, generating SSD predictions with the semantic segmentation model, from the target images, and producing an image of the area of interest, based on the SSD predictions, to identify a location of the resource.

[0012] According to another embodiment, there is a computing device for finding a sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource. The computing device includes an interface for receiving target images associated with an area of interest, and a processor connected to the interface. The processor is configured to generate pseudo-labels with a foundation model by prompting the foundation model with input images, which are different from the target images, train a semantic segmentation model based on the pseudo-labels, generate SSD predictions with the semantic segmentation model, from the target images, and produce an image of the area of interest, based on the SSD predictions, to identify a location of the resource.

[0013] According to yet another embodiment, there is a method for finding a sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource, and the method includes receiving target images associated with an area of interest, generating pseudo-labels with a foundation model byprompting the foundation model with annotated SSD, training a classification model based on the pseudo-labels to learn plural classes of SSDs, generating SSD prediction types with the classification model, from the target images, and producing an image of the area of interest, based on the SSD predictions, to identify a location of the resource.

[0014] According to still another embodiment, there is a computing device for finding a sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource, and the computing device includes an interface for receiving target images associated with an area of interest, and a processor connected to the interface and configured to generate pseudo-labels with a foundation model by prompting the foundation model with annotated SSD, train a classification model based on the pseudo-labels to learn plural classes of SSDs, generate SSD prediction types with the classification model, from the target images, and produce an image of the area of interest, based on the SSD predictions, to identify a location of the resource.

[0015] According to another embodiment, there is a method for finding a subcircular surface depression, SSD, which contributes to exploration of a subsurface resource, and the method includes receiving target images associated with an area of interest, generating less than 10 annotated SSDs from the target images, feeding the less than 10 annotated SSDs to a foundation model to generate SSD predictions, and producing an image of the area of interest, based on the SSD predictions, to identify a location of the resource.BRIEF DESCRIPTION OF THE DRAWINGS

[0016] For a more complete understanding of the present invention, reference is now made to the following descriptions taken in conjunction with the accompanying drawings, in which:

[0017] FIG. 1 is a flow chart of a traditional supervised machine learning method for detecting SSDs;

[0018] FIG. 2 is a flow chart of a method for using SSD-untrained foundation models for determining SSDs;

[0019] FIG. 3 is a flow chart of a method for processing input data prior to being provided to a given model for determining the SSDs;

[0020] FIG. 4 schematically illustrates a method that uses a few annotated SSDs and a foundation model, which is not SSD trained, for determining SSDs;

[0021] FIG. 5 schematically illustrates a method that uses a foundation model to produce labels to train a segmentation model;

[0022] FIG. 6 illustrates a prompt image with annotated SSDs;

[0023] FIG. 7A illustrates input data with possible SSDs and FIG. 7B illustrates output data, produced by the methods discussed above, that shows the predicted SSDs;

[0024] FIG. 8 schematically illustrates a method for determining the SSDs and identifying their types;

[0025] FIG. 9 is a flow chart of a method that uses two different machine learning models for determining the SSDs;

[0026] FIG. 10 is a flow chart of a method for determining the SSDs and identifying their types;

[0027] FIG. 11 is a flow chart of a method that uses annotated SSDs and a foundation model, which is not SSD trained, for determining SSDs; and

[0028] FIG. 12 is a schematic diagram of a computing device that is configured to implement any of the methods discussed herein.DETAILED DESCRIPTION OF THE INVENTION

[0029] The following description of the embodiments refers to the accompanying drawings. The same reference numbers in different drawings identify the same or similar elements. The following detailed description does not limit the invention. Instead, the scope of the invention is defined by the appended claims. The following embodiments are discussed with regard to the automatic detection and classification of SSDs, which might be associated with the location of a geologic feature or body in a subsurface, using terminology of gas data processing. However, the embodiments to be discussed next are not limited to gas data, but may be applied to other types of resources, for example, oil, mineral resources. Also, the embodiments discussed herein are not limited to automatic detection and classifying of SSDs, but may be used only for SSDs detection.

[0030] Reference throughout the specification to “one embodiment” or “an embodiment” means that a particular feature, structure or characteristic described in connection with an embodiment is included in at least one embodiment of the subject matter disclosed. Thus, the appearance of the phrases “in one embodiment” or “in an embodiment” in various places throughout the specification is not necessarily referring to the same embodiment. Further, the particular features, structures or characteristics may be combined in any suitable manner in one or more embodiments.

[0031] According to an embodiment, a versatile machine learning approach that allows the detection of SSDs in any given location is introduced. This approach relies on using pretrained (not SSD related) models without the need for training them on large manually annotated datasets. In other words, these models have beenpreviously trained for other purposes and they are readily available. This approach makes the overall process of detecting SSDs more efficient from a practical perspective. Foundation vision models, also known as large vision models, such as SAM

[0011] , SegGPT

[0012] , and DINOv

[0013] may be used for this approach. These are generalist vision models that are expected to perform well on a variety of vision tasks across different domains and can be adapted to specific applications with prompt engineering in a zero-shot

[0011] , one-shot or few-shot [12,13] manner. Thus, in this embodiment, it is possible to segment and map SSDs across an area by providing a right prompt (i.e., SSD input image) to the foundation model. A foundation model is understood herein to describe machine learning models trained on a broad spectrum of generalized data and capable of performing a wide variety of general tasks.

[0032] In this or another embodiment, it is possible to use the foundation model predictions as pseudo-labels to subsequently build a supervised learning dataset. Refining and augmenting pseudo-labels is much less tasking than manually creating annotations from scratch. These pseudo-labels are then used to train a semantic segmentation model (e.g., a different one from the ones discussed above), such as UNet

[0014] or ViT

[0015] . In one application, a machine learning classification model is used to classify SSDs segmented in pseudo-labels into their different categories such as lakes, permafrost, arctic sinkholes, and karstic depressions. Applicable classification models include convolutional neural network (CNN) based architectures as well as tree-based models.

[0033] Before discussing these methods in more detail, it is noted that an impact crater, sinkhole, volcanic calderas, or structural basins are sometimes calledSSD in the art. For example, the structural basins, which are formed in tectonicallyactive regions, may correspond to sub-circular depressions that can form as a result of crustal movements and subsidence. These formations are associated with faults or other geological processes and thus, may be indicative of oil and gas reservoirs. Thus, the methods discussed herein are implemented to predominantly find these formations.

[0034] According to an embodiment, which is illustrated in FIG. 2, a method for determining an SSD starts at step 200 by acquiring data, e.g., obtaining the data from a client and / or from publicly available databases. The data may include various types of data, as discussed later. This step may also pre-process the data, for example rescale one or more parts of the data so that data acquired from various sources may be stacked together. In step 202, the pre-processed data is used for generating pseudo-labels, e.g., the data may be fed along with few annotated data (e.g., less than 10 in one application) to prompt a foundation model for generating pseudo-labels. Note that the foundation model in this step may be any known model, which was previously trained for various purposes, unrelated to SSD. No specific training for the SSD is necessary in this step. However, the time and effort for performing this step is much smaller than what the industry is currently doing.

[0035] The annotated images (pseudo-labels) are then fed in step 204 to a semantic segmentation model for training. A segmentation model for image recognition is a type of neural network architecture designed to perform pixel-level categorization of an image. Unlike classification models that predict a single label for an entire image, segmentation models are designed to assign a label to each pixel, outlining distinct regions or objects within the image. A semantic segmentation modelis configured to assign each pixel of an input image to a specific class label, for example, SSD. Thus, this step trains the model for recognizing SSDs.

[0036] In step 206, an image of interest is fed to the trained semantic segmentation model for finding one or more SSDs, which are associated in step 208 with a hydrogen reservoir or accumulation or the presence of a hydrogen source rock or other hydrogen bearing mixtures. It is noted that the process of finding underground resources discussed above is much faster than a traditional process because the used model is not extensively SSD trained prior to finding the SSDs. In other words, a first general model that was never trained for SSDs may be used in step 202 for generating the pseudo-labels. This step is thus extremely fast. Contrary to this approach, the traditional methods spend a large amount of time to manually generate a large library of labels with specifically designed models.

[0037] A couple of possible implementations of the method illustrated in FIG. 2 are now discussed. First, the step 200 of acquiring and pre-processing the data, which might be common to all the methods, is discussed with regard to FIG. 3. In this example, satellite images (e.g., obtained from Sentinel 2 satellite) and also DEM data may be used. The DEM is derived from various sources (e.g., topographic maps, high-resolution light detection and ranging, etc.) and is a representation of the Earth's surface topography or terrain in digital form. It provides a representation of the Earth's surface by assigning elevations to geographical coordinates. Satellitebased DEMs are often generated using data collected by remote sensing instruments onboard Earth-observing satellites. These instruments may include radar sensors, stereo-pair optical sensors, or LiDAR (Light Detection and Ranging) systems. Other type of data may be used. Although the present embodiment showsthe use of two different types of data, those skilled in the art would understand that the embodiment also works with more than two types of data.

[0038] For example, a plurality of satellite images and DEM along with derivative images may be used as input. Freely available satellite and DEM datasets are used in this embodiment for world coverage to ensure that the approach can be applied globally. However, one skilled in the art would understand that commercial or client remote sensing data may be used in combination with or instead of the freely available data noted above. Upon selection of a target area of interest (AOI), the following imagery may be acquired / obtained to cover the AOI: (1) at least three Sentinel 2 L2A bands 310, e.g., 1 , 2, 3, 4, 5, 6, 7, 8A, 11 , and 12, which are resampled to a uniform 20m resolution (other resolutions may be selected), and (2) Copernicus 30m DEM 312. Those skilled in the art would understand that other types of images may be used.

[0039] Next, these images (or rasters) are homogenized and combined into a single raster. The Sentinel 2 raster 310 is reprojected to the same coordinate reference system (CRS) as the DEM in step 314 while the DEM 312 is resampled in step 316 to 20m resolution to match that of the Sentinel 2 image 310. Then, the areas of the two images that do not intersect are cropped out in step 318. The two images are combined in step 320 into a single raster 322 by stacking them along the channel dimension.

[0040] In one application, it is possible to generate derivative images from the Sentinel 2 bands, for example, soil-adjusted vegetation index (SAVI), enhanced vegetation index (EVI), ferrous iron, ferrous silicates ratio 1 , ferrous silicates ratio 2, etc. Other derivatives may also be computed from the DEM such as slope andcurvature. These derivatives can be calculated in real time and therefore, are not precomputed and stored. However, in one application, these derivates may be calculated before performing the method in FIG. 2. Note that any of these derivative images may be used instead of images 310 or even images 312. The step 200 discussed above may be common to different methods now discussed.

[0041] According to a first embodiment, FIG. 4 schematically illustrates a method for using prompt engineering to adapt an existing image processing model to a specific area of the globe to generate SSD predictions. The term “prompt engineering” is used herein to refer to any input (e.g., an image) that can be fed to the image processing model (i.e., a prompt-able foundation model) for triggering an image analysis for finding SSDs in the input image. More specifically, the proposed method relies on using a large pretrained model (e.g., SAM, SegGPT, or DINOv) 400 to map SSDs without having to build a large training dataset. For this method, the data is acquired and prepared as discussed above with regard to step 200. Then, one or more SSD annotations 402 are generated (e.g., up to 10). In this embodiment, a geologist manually annotated the SSDs 402. The annotated SSDs 402 were then rasterized 404 to transform the SSD polygons into a ground truth mask 406. Rasterizing image polygons to generate a ground truth mask in this context refers to the process of converting vector-based polygonal representations of objects or regions in the input image 401 (e.g., 310 and 312 discussed above with regard to FIG. 3) into pixel-based masks. The objects or regions of interest in an image are often defined using vector-based representations like polygons.Rasterization is the process of converting these vector representations into pixelbased images. In the context of image annotation, this involves filling the areadefined by the polygon with pixels. The pixels inside the polygon are usually assigned a value to indicate that they belong to the object, while pixels outside the polygon remain unchanged or are assigned a different value to indicate background. The resulting pixel-based image is called a ground truth mask. It serves as a reference or labeled data for training and evaluating machine learning models, particularly those designed for segmentation tasks. Each pixel in the mask corresponds to a specific location in the original image and indicates whether that location belongs to the object of interest or not.

[0042] Next, original images 401 (that describe the AOI, called herein “target images”) and the ground truth mask 406 are supplied 408 as prompts to the foundation model 400. Images 401 may include two different types of images 310 and 312, as previously discussed. This prompt engineering strategy includes using a single representative example (one-shot) or stacking multiple examples (few-shot) or cropping and stitching masks from different examples for memory efficiency or prompt learning algorithms (see, for example,

[0016] ). The model 400 then generates SSD predictions 410 over the desired AOI. Then, in step 412, the SSD predictions 410 can be associated with an underground reservoir / accumulation / source rock and various exploration activities may be performed for sampling and exploitation. In one application, an image is generated in this step that indicates an area(s) for further investigation. In one application, it is possible to sample the soil inside and around the SSD for confirming the presence of the reservoir / accumulation / source rock. Alternatively, it is possible to use further imaging techniques to explore for the presence of the reservoir / accumulation / source rock. Such imaging techniques maybe seismic acquisition. Other procedures may be used for confirming the presence of the reservoir / accumulation / source rock.

[0043] According to a second embodiment, which is illustrated in FIG. 5, two different machine learning models are used for generating the SSD predictions. A first model is the foundation model 400 discussed above and the second model is a semantic segmentation model 500. In one application, the two models are identical. However, for expediting the SSD prediction process, the second model 500 is selected to be smaller than the first model 400, i.e., the second model has less hidden layers than the first model or is mathematically less complex than the first model. In this way, the second model is much faster in processing the input data 401 than the first model.

[0044] The first model 400 is fed in this embodiment with input images 502 that may not cover the AOI. These images may be any of the images discussed above with regard to FIG. 3. Note that the first model 400 has not been trained to recognize SSD. The predictions of the foundation model 400 are considered to be pseudo-labels 504. Because the input images 502 are not images from the AOI, the pseudo-labels 504 are not restricted to the AOI, i.e., they may be indicative of SSD from various parts of the globe. The pseudo-labels may be rasterized to generate a ground truth mask as noted above. In one application, the pseudo-labels 504 may be refined and augmented by adding any missing SSDs and removing false positives. This step may be performed by a geologist and the input images may be images in which SSDs are known to exist (from a region or regions of the globe that are not similar to the AOI). In this way, a training dataset 506 is obtained without the need toannotate it from scratch. This means that a large amount of time is saved as it is known that the annotation step is the most time consuming for the existing methods.

[0045] This dataset 506 is used to train the second model 500, for example, a semantic segmentation model. In one implementation, the following steps are used for training the model. The refined pseudo-labels are rasterised into a binary mask as discussed above with regard to steps 404 and 406. Using a sliding window, the image and the mask are concurrently chipped, i.e., broken into smaller parts. The resulting image / mask chip pairs are then split into training and validation sets which are used to train and evaluate the second model in step 204. Image augmentation is used during training to increase the variety of training samples and regularize the model. The augmentation technique may include any one of or a combination of random horizontal and vertical flipping, transpose, rotation, or random cropping. After this step, the optimized SSD segmentation model 508 is used to generate SSD predictions 510, similar to the predictions 410. In step 512, the SSD predictions 510 may be associated with an underground reservoir / accumulation / source rock and various exploration activities may be performed for sampling and exploitation. As noted above, an image may be generated in this step and this image indicates an area(s) for further investigation.

[0046] The methods discussed above with regard to FIGs. 4 and 5 are capable to determine whether a region in a target image is an SSD region or not. For example, FIG. 6 shows an image 310 (target image), with annotated SSDs 402 or predicted SSDs 504 (pseudo-labels). Note that a shape of the SSD does not have to be circular. FIG. 7A shows an input image 401 / 310 that is used in the method of FIG. 5 for determining the location of the SSDs. Two potential SSDs 710 are illustrated inthe figure. FIG. 7B shows the output 510 of the second model 500 for FIG. 5. It is noted that the output 510, i.e., the SSD prediction areas, correspond to the actual SSDs 710. However, these methods do not produce any indication about what type of SSD is found in the target images.

[0047] To address this matter, a different method is now discussed with regard to FIG. 8. The steps already discussed above are not explained again. This method is configured to classify the SSDs into different types, for example, fairy circles, lakes, permafrost, and sinkholes. Other categories or types of SSDs may be added to the method. In this method, the foundation model 400 is used to generate the pseudo-labels 506. The input images 401 are cropped in step 820 around the SSD pseudo-labels 506 and a class label is assigned in step 822, to the cropped images, regarding the SSD type, to generate a training dataset. This dataset is used to train a classification model 800 to obtain an optimized SSD classification model 826. Any type of classification model may be used. The output of the optimized classification model 826 are used in step 810, which is similar to step 510, to make SSD predictions. These predictions include, in addition to the location of the SSD, its type. Then, in step 812, which is similar to step 512, the SSD predictions may be used to identify an area(s) for further investigation. Classifying SSDs into different types or categories is more insightful than simple binary segmentation and makes the approaches discussed herein suitable for a wide range of applications beyond native hydrogen exploration, e.g., permafrost monitoring. This method may be combined with the methods illustrated in FIGs. 4 and 5.

[0048] Based on the embodiments discussed above, various methods for generating SSD predictions for identifying a location of a resource in the subsurfaceare now discussed with regard to FIGs. 9 to 11 . A first method, illustrated in FIG. 9, includes a step 900 of receiving target images 401 associated with an area of interest, a step 902 of generating pseudo-labels 506 with a foundation model 400 by prompting the foundation model 400 with input images 502, which are different from the target images 401 , a step 904 of training a semantic segmentation model 500 based on the SSD predictions 504, a step 806 of generating SSD predictions 510 with the semantic segmentation model 500, from the target images 401 , and a step 908 of producing an image of the area of interest, based on the SSD predictions 510, to identify a location of the resource.

[0049] In one application, the foundation model is not pre-trained with SSD based images for generating the pseudo-labels and the semantic segmentation model is less complex than the foundation model. The method may further include generating a training dataset from the pseudo-labels by adding missing SSDs, and using the training dataset to train the semantic segmentation model. In one application, the target images include a first type of images 310 and a second type of images 312, which is different from the first type. The first type of images are one of Sentinel 2 L2A bands images, or derivative images of the Sentinel 2 L2A bands, and the second type of images are digital elevation model images. The derivative images include any of soil-adjusted vegetation index, enhanced vegetation index, ferrous iron, ferrous silicates ratio 1 , or ferrous silicates ratio 2.

[0050] In another embodiment, as illustrated in FIG. 10, there is a method for finding a subsurface resource based on a sub-circular surface depression, SSD, and the method includes a step 1000 of receiving target images 401 associated with an area of interest, a step 1002 of generating pseudo-labels 506 with a foundationmodel 400 by prompting the foundation model 400 with annotated SSD 402, a step 1004 of training a classification model 800 based on the annotated SSD 402 to learn plural classes of SSDs, a step 1006 of generating SSD prediction types 810 with the classification model 800, from the target images 401 , and a step 1008 of producing an image of the area of interest, based on the SSD predictions 410, to identify a location of the resource.

[0051] In one application, the SSD prediction types include fairy circles, lakes, permafrost, and sinkholes. The foundation model is not pre-trained with SSD based images for generating the pseudo-labels. The method may further include generating a training dataset from the pseudo-labels by adding missing SSDs, and using the training dataset to train the classification model. The target images include a first type of images 310 and a second type of images 312, which is different from the first type. The first type of images are one of Sentinel 2 L2A bands images, or derivative images of the Sentinel 2 L2A bands, and the second type of images are digital elevation model images. The derivative images include one of soil-adjusted vegetation index, enhanced vegetation index, ferrous iron, ferrous silicates ratio 1 , or ferrous silicates ratio 2.

[0052] According to yet another embodiment, as illustrated in FIG. 11 , there is a method for finding a subsurface resource based on a sub-circular surface depression, SSD, and the method includes a step 1100 of receiving target images 401 associated with an area of interest, a step 1102 of generating less than 10 annotated SSDs 402 from the target images 401 , a step 1104 of feeding the less than 10 annotated SSDs 402 to a foundation model 400 to generate SSD predictions410, and a step 1106 of producing an image of the area of interest, based on the SSD predictions 410, to identify a location of the resource.

[0053] The methods discussed herein may be applied to the field of subsurface exploration, for example, hydrogen gas exploration, hydrocarbon exploration and development, geothermal exploration and development, and carbon capture and sequestration, or other natural resource exploration and exploitation.

[0054] The above-discussed procedures and methods may be implemented in a computing device as illustrated in FIG. 12. Hardware, firmware, software, or a combination thereof may be used to perform the various steps and operations described herein. The computing device 1200 is suitable for performing the activities described in the above embodiments and may include a server 1201 . Such a server 1201 may include a central processor (CPU) 1202 coupled to a random access memory (RAM) 1204 and to a read-only memory (ROM) 1206. ROM 1206 may also be other types of storage media to store programs, such as programmable ROM (PROM), erasable PROM (EPROM), etc. Processor 1202 may communicate with other internal and external components through input / output (I / O) circuitry 1208 and bussing 1210 to provide control signals and the like. Processor 1202 carries out a variety of functions as are known in the art, as dictated by software and / or firmware instructions. In one application, one or more graphic process units (GPUs) may be added to accelerate computation of the images discussed above.

[0055] Server 1201 may also include one or more data storage devices, including hard drives 1212, CD-ROM drives 1214 and other hardware capable of reading and / or storing information, such as DVD, etc. In one embodiment, software for carrying out the above-discussed steps may be stored and distributed on a CD-ROM or DVD 1216, a USB storage device 1218 or other form of media capable of portably storing information. These storage media may be inserted into, and read by, devices such as CD-ROM drive 1214, disk drive 1212, etc. Server 1201 may be coupled to a display 1220, which may be any type of known display or presentation screen, such as LCD, plasma display, cathode ray tube (CRT), etc. A user input interface 1222 is provided, including one or more user interface mechanisms such as a mouse, keyboard, microphone, touchpad, touch screen, voice-recognition system, etc. Server 1201 may be coupled to other devices, such as satellite data repository. The server may be part of a larger network configuration as in a global area network (GAN) such as the Internet 1228, which allows ultimate connection to various landline and / or mobile computing devices.

[0056] As described above, the apparatus 1200 may be embodied by a computing device. However, in some embodiments, the apparatus may be embodied as a chip or chip set. In other words, the apparatus may comprise one or more physical packages (e.g., chips) including materials, components and / or wires on a structural assembly (e.g., a baseboard). The structural assembly may provide physical strength, conservation of size, and / or limitation of electrical interaction for component circuitry included thereon. The apparatus may therefore, in some cases, be configured to implement an embodiment of the present invention on a single chip or as a single “system on a chip.” As such, in some cases, a chip or chipset may constitute means for performing one or more operations for providing the functionalities described herein.

[0057] In an example embodiment, the processor 1202 may be configured to execute instructions stored in the memory device 1204 or otherwise accessible tothe processor. Alternatively, or additionally, the processor may be configured to execute hard coded functionality. As such, whether configured by hardware or software methods, or by a combination thereof, the processor may represent an entity (e.g., physically embodied in circuitry) capable of performing operations according to an embodiment of the present invention while configured accordingly. Thus, for example, when the processor is embodied as an ASIC, FPGA or the like, the processor may be specifically configured hardware for conducting the operations described herein. Alternatively, as another example, when the processor is embodied as an executor of software instructions, the instructions may specifically configure the processor to perform the algorithms and / or operations described herein when the instructions are executed. However, in some cases, the processor may be a processor of a specific device (e.g., a pass-through display or a mobile terminal) configured to employ an embodiment of the present invention by further configuration of the processor by instructions for performing the algorithms and / or operations described herein. The processor may include, among other things, a clock, an arithmetic logic unit (ALU) and logic gates configured to support operation of the processor.

[0058] The term “about” is used in this application to mean a variation of up to 20% of the parameter characterized by this term. It will be understood that, although the terms first, second, etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first object or step could be termed a second object or step, and, similarly, a second object or step could be termed a first object or step, without departing from the scope of the presentdisclosure. The first object or step, and the second object or step, are both, objects or steps, respectively, but they are not to be considered the same object or step.

[0059] The terminology used in the description herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used in this description and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any possible combinations of one or more of the associated listed items. It will be further understood that the terms "includes," "including," "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. Further, as used herein, the term "if" may be construed to mean "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context.

[0060] The disclosed embodiments provide an automatic detection of SSDs, which may be associated with a resource reservoir / accumulation / source rock in the subsurface. It should be understood that this description is not intended to limit the invention. On the contrary, the embodiments are intended to cover alternatives, modifications and equivalents, which are included in the spirit and scope of the invention as defined by the appended claims. Further, in the detailed description of the embodiments, numerous specific details are set forth in order to provide a comprehensive understanding of the claimed invention. However, one skilled in theart would understand that various embodiments may be practiced without such specific details.

[0061] Although the features and elements of the present embodiments are described in the embodiments in particular combinations, each feature or element can be used alone without the other features and elements of the embodiments or in various combinations with or without other features and elements disclosed herein.

[0062] This written description uses examples of the subject matter disclosed to enable any person skilled in the art to practice the same, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the subject matter is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to be within the scope of the claims.ReferencesThe entire content of all the publications listed herein is incorporated by reference in this patent application.[1] N. Juergens, “The Biological Underpinnings of Namib Desert Fairy Circles,” Science, 2013, vol. 339, pp.1618-1621 .[2] B. Malvoisin and F. Brunet, “Barren ground depressions, natural H2 and orogenic gold deposits: Spatial link and geochemical model,” Science of the Total Environment, 2023, vol. 856, p.158969.[3] M.D. Cramer and N.N. Barger, “Are Namibian “fairy circles” the Consequence of Self-Organizing Spatial Vegetation Patterning?” PloS one, 2013, vol. 8, p.e70876.[4] S. Getzin, H. Yizhaq and W.R. Tschinkel, “Definition of “fairy circles” and how they differ from other common vegetation gaps and plant rings.” Journal of Vegetation Science, 2021 , vol 32, p.e13092.[5] E. Frery, L. Langhi, M. Maison and I. Moretti, “Natural hydrogen seeps identified in the North Perth Basin, Western Australia.” International Journal of Hydrogen Energy, 2021 , vol. 46, pp.31158-31173.[6] M. Lundine and A. C. Trembanis, “Developing a CNN for automated detection of Carolina bays from publicly available LiDAR data,” in AGU Fall Meeting Abstracts, 2019, vol. 2019, pp. EP54C-03.[7] M. Lundine and A. Trembanis, “Using Convolutional Neural Networks for Detection and Morphometric Analysis of Carolina Bays from Publicly Available Digital Elevation Models,” Remote Sensing, vol. 13, no. 18, p. 3770, Sep. 2021 , doi:10.3390 / rs13183770.[8] W. J. Leong, B. Yadav, A. Negrete, I. M. Howat, and J. Moortgat, “H2OvalNet: Detecting Fairy Circles from Sentinel-2 Imagery in Australia,” in AGU Fall Meeting Abstracts, 2022, vol. 2022, pp. H22P-1039.[9] Y. Zhu, Z. Moayed, B. Bollard-Breen, A. Doshi, J. B. Ramond and R. Klette, “Detection of Fairy Circles in UAV Images Using Deep Learning,” 2018 15th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), Auckland, New Zealand, 2018, pp. 1-6, doi: 10.1109 / AVSS.2018.8639450.

[0010] K. Noy, M. Silver, O. Pesek, H. Yizhaq, E. Marais, and A. Karnieli, “Spatial and spectral analysis of fairy circles in Namibia on a landscape scale using satellite image processing and machine learning analysis,” International Journal of Applied Earth Observation and Geoinformation, vol. 121 , p. 103377, 2023.

[0011] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. “Segment anything,” arXiv preprint arXiv:2304.02643, 2023.

[0012] X. Wang, X. Zhang, Y. Cao, W. Wang, C. Shen, and T. Huang, “Seggpt: Segmenting everything in context,” arXiv preprint arXiv:2304.03284, 2023.

[0013] F. Li et al., “Visual In-Context Prompting.” arXiv, 2023. doi: 10.48550 / ARXIV.2311 .13601 .

[0014] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, 2015, pp. 234-241.

[0015] A. Dosovitskiy et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929, 2020.

[0016] K. Chen et al., “RSPrompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model,” arXiv preprint arXiv:2306.16269,2023.

Claims

WHAT IS CLAIMED IS:1 . A method for finding sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource, the method comprising: receiving (900) target images (401) associated with an area of interest; generating (902) pseudo-labels (506) with a foundation model (400) by prompting the foundation model (400) with input images (502), which are different from the target images (401); training (904) a semantic segmentation model (500) based on the pseudolabels (506); generating (906) SSD predictions (510) with the semantic segmentation model (500), from the target images (401); and producing (908) an image of the area of interest, based on the SSD predictions (510), to identify a location of the resource.

2. The method of Claim 1 , wherein the foundation model is not pre-trained with SSD based images for generating the pseudo-labels.

3. The method of Claim 1 , wherein the semantic segmentation model is less complex than the foundation model.

4. The method of Claim 1 , further comprising: generating a training dataset from the pseudo-labels by adding missing SSDs; andusing the training dataset to train the semantic segmentation model.

5. The method of Claim 1 , wherein the target images include a first type of images (310) and a second type of images (312), which is different from the first type.

6. The method of Claim 5, wherein the first type of images include one of Sentinel 2 L2A bands images, or derivative images of the Sentinel 2 L2A bands, and the second type of images include digital elevation model images.

7. The method of Claim 6, wherein the derivative images include one of soil- adjusted vegetation index, enhanced vegetation index, ferrous iron, ferrous silicates ratio 1 , or ferrous silicates ratio 2.

8. A computing device (1200) for finding a sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource, the computing device (1200) comprising: an interface (1208) for receiving (900) target images (401) associated with an area of interest; and a processor (1202) connected to the interface (1208) and configured to, generate (902) pseudo-labels (506) with a foundation model (400) by prompting the foundation model (400) with input images (502), which are different from the target images (401);train (904) a semantic segmentation model (500) based on the pseudo-labels(506); generate (906) SSD predictions (510) with the semantic segmentation model (500), from the target images (401); and produce (908) an image of the area of interest, based on the SSD predictions (510), to identify a location of the resource.

9. A method for finding a sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource, the method comprising: receiving (1000) target images (401) associated with an area of interest; generating (1002) pseudo-labels (506) with a foundation model (400) by prompting the foundation model (400) with annotated SSD (402); training (1004) a classification model (800) based on the pseudo-labels (506) to learn plural classes of SSDs; generating (1006) SSD prediction types (810) with the classification model (800), from the target images (401); and producing (1008) an image of the area of interest, based on the SSD predictions (410), to identify a location of the resource.

10. The method of Claim 9, wherein the SSD prediction types include fairy circles, lakes, permafrost, and sinkholes.11 . The method of Claim 9, wherein the foundation model is not pre-trained with SSD based images for generating the pseudo-labels.

12. The method of Claim 9, further comprising: generating a training dataset from the pseudo-labels by adding missing SSDs; and using the training dataset to train the classification model.

13. The method of Claim 9, wherein the target images include a first type of images (310) and a second type of images (312), which is different from the first type.

14. The method of Claim 13, wherein the first type of images are one of Sentinel 2 L2A bands images, or derivative images of the Sentinel 2 L2A bands, and the second type of images are digital elevation model images.

15. The method of Claim 14, wherein the derivative images include one of soil-adjusted vegetation index, enhanced vegetation index, ferrous iron, ferrous silicates ratio 1 , or ferrous silicates ratio 2.

16. A computing device (1200) for finding a sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource, the computing device (1200) comprising: an interface (1208) for receiving (1000) target images (401) associated with an area of interest; and a processor (1202) connected to the interface (1208) and configured to,generate (1002) pseudo-labels (506) with a foundation model (400) by prompting the foundation model (400) with annotated SSD (402); train (1004) a classification model (800) based on the pseudo-labels (506) to learn plural classes of SSDs; generate (1006) SSD prediction types (810) with the classification model(800), from the target images (401); and produce (1008) an image of the area of interest, based on the SSD predictions (410), to identify a location of the resource.

17. A method for finding a sub-circular surface depression, SSD, which contributes to exploration of a subsurface resource, the method comprising: receiving (1100) target images (401) associated with an area of interest; generating (1102) less than 10 annotated SSDs (402) from the target images (401); feeding (1104) the less than 10 annotated SSDs (402) to a foundation model(400) to generate SSD predictions (410); and producing (1106) an image of the area of interest, based on the SSD predictions (410), to identify a location of the resource.

18. The method of Claim 17, wherein the foundation model is not pre-trained with SSD based images for generating the SSD predictions.

19. The method of Claim 17, wherein the target images include a first type of images (310) and a second type of images (312), which is different from the first type.

20. The method of Claim 19, wherein the first type of images are one of Sentinel 2 L2A bands images, or derivative images of the Sentinel 2 L2A bands, and the second type of images are digital elevation model images.

Images