This science case is focused on the detection and mapping of sinkhole-like depressions on Mars and the Moon through Deep Learning Object Detection and Instance Segmentation.
The term sinkhole refers to different morphologies that have a in common the processes of depleting materials of different type into an area within the morphology itself (Waltham, 2005). On Earth the formation of sinkholes is related to a cluster of processes, and could occur in various type of grounds, furthermore presence of water has a key-role. On other terrestrial planets, although the mechanisms for the origin of these landforms are similar, if not the same, with the main difference that as far as we know there is no liquid water that can be involved in the formation of these landforms, and therefore the mechanisms and processes are still debated. Several authors suggested the hypothesis of formation from lava tube collapses (Greeley, 1971; Cruikshank and Wood, 1972; Carr et al., 1977), others imply different volcanic and tectonic processes involved (Wyrick et al., 2004). In karst environment, sinkholes are called “doline” and are more related to processes connected to dissolution of carbonates and evaporites, while in volcanic environments they are called “pit craters” and are more related to process connected to lava tube stability.
Doline, pit craters, pit chains and lava tubes are well-known morphologies on Earth (Lauterbach et al., 2019; Díaz Michelena et al., 2020), Mars (Carr, 1977; Cushing et al., 2007; Cushing, 2012; Cushing et al., 2015), Moon (Chappaz et al., 2017), Venus (Sawford et al., 2015), Mercury (Gillis-Davis et al., 2009), and Ganymede and Saturnian satellites (Barlow et al., 2017). Studies suggested the presence of sinkholes both in karst (Baioni and Tramontana, 2015, 2016) and volcanic terrains of Mars (Sauro et al., 2020), in particular thanks to HiRISE and CTX cameras on board MRO more than 1000 potential cave entrances have been identified on Mars (Cushing, 2012), while on Moon thanks to SELENE and LRO cameras more than 300 (Wagner and Robinson, 2014).
In the framework of geological exploration of terrestrial planets like Earth, those landforms - being a potential direct access to subsurface - are one of the most promising environments where to focus the research of valuable data of different kind, from planet's geological stratigraphy (Lauterbach et al., 2019; Sauro et al., 2020a) to valuable ore deposits (Blamont, 2014). They could also provide access to cave entrances where it is possible to gather even more data since caves are a natural shelter from cosmic radiation, and thus potentially provide a feasible environment for searching of life traces (Léveillé and Datta, 2010; Cushing, 2012), for the development of future human outposts (Cushing, 2012; Blamont, 2014; Carrer et al., 2018), and for the planning of future missions (Hare et al., 2018).
Detecting, mapping, and describing sinkhole-like landform is a challenging process since a set of tedious tasks must be conducted manually, from data collection to manual analysis, mapping using Geographic Information Systems (GIS) software and extracting morphometric parameters.
For Mars, there exists a downloadable database of more than 1000 cave candidates (Cushing, 2012) and there are several other publications that analyse possible cave entrances on Mars and Moon (Cushing et al., 2007; Sauro et al., 2020). Also, for the Moon there is an online database, but it is not downloadable. Both of the databases are region specific or with preferred planetary spots, are created with manual or semi-automated methods and are not updated that regularly. Starting from literature, five classes have been identified:
As many of these landforms are extremely similar, a crucial role is the labelling of all classes and the creation of a modest initial data set to be used for training, validating and testing.
There are different challenges in this science case:
The minimal results are images with a superimposed bounding box with class and score for each detection. The images still need to be further processed manually.
A perfect result is a GIS geopackage containing a polygon-type layer with all of the detections, the name of the image and the score.
Mars optical images are provided by the Martian Reconnaissance Orbiter (MRO), and are composed of:
Lunar optical images are provided by the LRO spacecraft, and are composed of:
All data can be accessed through NASA PDS Geoscience Node ODE Portal.
Considering the five classes that have been identified and taking into account their similarities within each other, the problem could be described as a multi-label multi-class object detection problem. The data consist of all images, the best suitable approach is Deep Learning since it is widely applied on Earth for similar problems and could be adapted easily to different targets.
Several Deep Learning object detection algorithms exits, in addition to open source framework based on Python and C/C++ programming language.
Without entering in the depth of Deep Learning, abstraction problem and Convolutional Neural Networks architecture, these algorithm can be described as an algorithm that take input images, attempt to extract the essential features that describe these images by using convolution layers, pooling layers and fully connected layers, tries to predict unseen images and compare the obtained results with ground truth data to check whether continue the training or go back and change training parameters.
The model resulting from the training phase could be used to inference newer images.
Training results are strictly related to data set dimension and labeling quality, and this is the major problematic and task to be accounted since in this science case there are huge base data sets to be used but almost no information on which specific image to be used, and no labels available.
As a first step, a deep learning object detection algorithm YOLO (DeepLandforms-YOLOv5) has been applied to detect pit craters on the surface of Mars from images obtained by the HiRISE instrument on board the MRO. Current performance reaches 74% recall and 70% precision. Image augmentation and expanded labelling were shown to provide better performance results. The extraction of statistically relevant products was also demonstrated.
To improve the results obtained by this first tool for automated mapping of pits, a change of architecture was necessary. The results obtained by that tool, despite their high quality, need further processing since they are not immediately usable for proper mapping as they are composed only by a pair of coordinates that localize the centre of the detected features. Such detections still need to be properly mapped as polygonal shapes by users.
Since this is a highly time-consuming and tedious task, it led to the development of a new tool based on Deep Learning Instance Segmentation, to retrieve not only point coordinates of the detected features, but also a polygonal shape. The obtained results were then compared to the results obtained with the previous tool and with the MGC^3 database (Cushing et al. 2012, 2015), showing good results. A publication and this new tool will be released soon.
Results of this science case were presented at the LPSC2021.
References: