Automatic Mapping of Small Lunar Impact Craters Using LRO-NAC Images

Impact craters are the most common feature on the Moon's surface. Crater size-frequency distributions provide critical insight into the timing of geological events, surface erosion rates, and impact fluxes. The impact crater size-frequency follows a power law (meter-sized craters are a few orders of magnitude more numerous than kilometric ones), making it tedious to manually measure all the craters within an area to the smallest sizes. We can bridge this gap by using a machine learning algorithm. We adapted a Crater Detection Algorithm to work on the highest resolution lunar image data set (Lunar Reconnaissance Orbiter-Narrow-Angle Camera [NAC] images). We describe the retraining and application of the detection model to preprocessed NAC images and discussed the accuracy of the resulting crater detections. We evaluated the model by assessing the results across six NAC images, each covering a different lunar area at differing lighting conditions. We present the model's average true positive rate for small impact craters (down to 20 m in diameter) is 93%. The model does display a 15% overestimation in calculated crater diameters. The presented crater detection model shows acceptable performance on NAC images with incidence angles ranging between similar to 50 degrees and similar to 70 degrees and can be applied to many lunar sites independent to morphology. Plain Language Summary The Moon's surface is covered in impact craters and recording their spatial density gives researchers the ability to study the geological evolution of our satellite. Analyzing craters helps in determining the physical properties of planetary surfaces and how/if impact rates change over time. These analyses rely on recording spatial densities for numerous surfaces, which has been achieved for craters > 1-2 km on the Moon. Manually counting the smaller craters, which number in the hundreds of millions, is a daunting task. We adapted a Crater Detection Algorithm and applied it to the highest resolution lunar imagery data set. We describe our method for gathering, reformatting, and detecting craters across lunar images down to 20 m in diameter. The detection model performance was quantitatively evaluated across six different regions, each with different terrain and lighting conditions. Comparison between manually mapped craters and detections from our model allows us to conclude that the model has an acceptable performance in detecting fresh to moderately degraded craters of all sizes, down to 20 m in diameter, when compared to other studies. Automated crater detection complements manual counting methods and aids in unlocking secrets of the Moon's surface.

Automated summary

Impact craters are the most common feature on the Moon's surface. A machine learning algorithm has been used to detect small impact craters with a 93% accuracy rate on high-resolution lunar images.