Catching "spiders" on Mars – investigation of Martian araneiform terrain by AI

• 1. Introduction

Fig. 1 Spider-like features or “spiders” on Mars, observed by remote sensing images

Atmosphere of Mars consists primarily of CO₂ (96% by volume) [1, 2], up to 25% atmosphere undergoes a seasonal CO₂ cycle of sublimation/condensation [1, 2]: in winter/autumn, atmospheric CO₂ condenses as seasonal CO₂ ice at the polar surface; in spring/summer, it sublimates back into the atmosphere. This seasonal CO₂ cycle creates many exotic phenomena, e.g., dark spots and araneiform terrain or “spiders” (Fig. 1) [3-9]. Their formation process is believed as [3-9]: solar energy penetrates seasonal CO₂ ice, later causing CO₂ gas jetting and dust eruption onto the ice surface, forming dark spots. With yearly repetitions, spiders are formed showing radial or dendritic troughs. Spider formation is directly linked to seasonal CO₂ cycling, thus Mars climate and an element in shaping the polar surface [10].

It is openly debated how climate affects spiders’ spatial configurations, geomorphologies and formation. To address these, sufficient spider observations are required. Currently the only spider observations are through vast remote sensing images as they have no earthly analogies. But this poses the following issue: how to effectively detect, then map spiders besides recruiting citizen scientists which possibly omits some spider observations due to understanding levels of spider formation.

This work focuses on applying super-detection capabilities of artificial intelligence (AI) for faster and more efficient detection of potential growth, spatial configuration and various morphologies of spiders based on remote sensing images, reducing uncertainties and low speed of human’s visual inspections which can be subjective, deepen our understanding of mechanisms behind spider formation.

• 2. Data and AI methods

We used HiRISE 5 Martian years images [11] at this stage. We tried to include an equal number of images for each type of spiders to prevent model bias.

Two types of AI methods were used in this work:

2.1 Conventional image processing techniques involve basic image manipulation such as image preprocessing to enhance contrasts. It uses local window binarization to highlight more intense pixels.

2.2 Deep learning with a single-stage detector applies “You Only Look Once” (YOLO, by Ultralytics) algorithm for object detection, which is more sophisticated suited for complex image recognition. To refine and enhance performance of YOLO detectors for spiders, several strategies are implemented, focusing on dataset improvement, model tuning, and advanced training techniques [12, 13].

a. Model Architecture and Parameter Tuning

Experiment with various YOLO Configurations, e.g., YOLOv8-Nano, Small, Medium and XLarge, help find the optimal balance between speed and accuracy for the detecting spiders task.

Hyperparameter Optimization are obtained by systematic testing of different hyperparameters, including learning rate, batch size, and number of epochs, to identify the best settings for the model.

b. Advanced Training Techniques

Transfer Learning is leveraged by starting with a model pre-trained on similar tasks and fine-tune it on the spider detection task. This significantly enhances the learning efficiency and final model performance.

Data Augmentation is expanded by incorporating random rotations, scaling, and other transformations to make the model robust to various spatial variations.

Multi-scale Training with images at multiple scales are to improve ability to detect various spiders.

• 3. Initial results and dissussion

Fig. 2 Results and confusion matrices for YOLO v8 size=nano (top), small (middle) and medium (bottom), with 50 (a and b) and 100 (c and d) epochs

Fig. 3 Predicted detection from the tunned deep learning model (left, Labels; right, predictions)

The method 1 had limited success in accurately detecting spiders due to the similar coloration of the spiders and their backgrounds and their complex, ramified structures. It is more effective for identifying dark spots, which are relatively distinct from their surroundings.

Figs. 2 and 3 show the method 2, the YOLO demonstrated superior performance in detecting spiders, benefiting from its ability to learn from the data and distinguish spiders from complex backgrounds. (1) Different configurations of YOLO (Fig. 2) were tested, with the smaller models (Nano, Small) providing a good balance between speed and accuracy, sufficient for the dataset used. (2) The model was effective even when spiders were not distinctly separable from the background, showing potential for identifying overlapping spiders.

The models were trained using various computational resources to optimize performance. Our results showed that extending training beyond 50 epochs did not significantly improve performance, likely due to the small size of dataset. Precision and recall were used as metrics, and the models achieved good results with relatively few false positives.

• 4. Conslusion

Primarily, we tested YOLO on the given images. Next, we plan to extend the implementation to more state-of-the-art AI approaches to evaluate the performances across multiple benchmarks. Such as, alternative deep learning models or more complex YOLO configurations which  can be used to enhance detection accuracy. Overall, the deep learning approach, particularly YOLO, proved to be a promising method for the automated detection of Martian spiders, indicating a significant step forward in the application of AI technologies for planetary sciences.

References. [1]Leighton, Murray, 1966. Science 153. [2]Nier, McElroy, 1977. JGR 82. [3]Kieffer, 2007. JGR 112. [4] Piqueux et al. 2003. JGR 108 (E8). [5]Portyankina et al. 2017. Icarus 282. [6]Hansen et al. 2010. Icarus 205. [7] Hao et al. 2019. Icarus 317. [8] Diniega et al. 2021. Geomorphology 380. [9]Mc Keown et al. 2021. Sci Rep 11. [10]Piqueux, Christensen, 2008. JGR 113. [11] McEwen et al. 2007. JGR 112. [12] Hasan et al. 2021. Comput. Electron. Agric. 184. [13]Lodhi et al. 2022. ICRITO 2022 1-6