EVA – Towards a Comprehensive Inventory of Small Volcanic Features on Venus

Current models of Venus’ volcanic activity and evolution largely rely on maps of volcanic edifices and features, most notably on an extensive human survey of Magellan data that produced a catalog of ~85,000 volcanic features with sizes ranging from ~1 km (~7 SAR pixels) to ~100 km [1]. Our goal is to use a supervised deep learning-driven approach to address some of the limitations of the catalog provided by [1], specifically by 1) suppressing or removing (human) observer expectancy and fatigue effects, 2) complementing the catalog with edifices smaller than ~5 km (making up 99 % of the [1] catalog), and 3) adding accurate morphometric information for all volcanic features. Ultimately, we seek to create a reference catalog of volcanic features on Venus that can enable investigations of volcanic activity and evolution at an unprecedented level of scale as well as form the backbone for the systematic identification of surface change using new data returned by upcoming Venus missions.

We previously trained EVA, the Extractor Vulcanis Aedificiis (Volcanic Edifice Extractor) [2]. The current version of EVA identifies ~200,000 shield volcanoes (shields) and calderas across Venus, scanning through both left-look (LL) and right-look (RL) SAR data in only ~4 hours. EVA detects ~90 % of all surface-visible edifices in a small, dedicated testset, where ~90 % of all testset detections are correct.  Globally, we find that the number of detected shields increases by up to ~5x in regions mapped as shield plains [3] and more than 10x in more heavily tectonized regions. Qualitative inspection suggests that in tectonized regions, false detections are more prevalent. To further scrutinize EVA’s performance, we examine 22 regions across the planet covered by LL data, each with at least ~50 EVA detections, that span all major geologic units. Nine regions are also covered by RL images.  For each region we manually map shields and calderas to provide a second “ground truth” data set (in addition to that of [1]) against which EVA detections can be compared. Preliminary results indicate that in regions of higher shield density [1], the two ground truth data sets differ in number of detections by a factor of 1.5-2, with excellent overlap (i.e. most features in the smaller data set are contained in the larger one), and EVA results in another ~2x as many detections. There is also very good agreement between the two ground-truth data sets in more tectonized areas, and we are currently using results from these areas to understand and refine EVA performance.  

We will present EVA’s current performance and solicit community feedback, to ensure that EVA can be a reliable, well-understood, and comprehensive inventory of volcanic features on Venus.

 

[1] Hahn & Byrne (2023). A Morphological and Spatial Analysis of Volcanoes on Venus. JGR Planets 128 (4).
[2] Bickel et al. (2025). Revisiting Volcanism on Venus with Deep Learning. Lunar and Planetary Science Conference 2025, Abstract ID #1387.
[3] Ivanov & Head (2011). Global Geological Map of Venus. Planetary and Space Science 59 (1559–1600).