Analysis Modules
Domain-specific AI/ML analysis modules that run inside BisQue.
BisQue supports a growing collection of analysis modules that bring purpose-built AI/ML workflows directly into the cloud platform. Each module opens its native scientific imagery, runs detection or reconstruction in the browser, and keeps results and annotations attached to the source data. The modules below span wildlife and ecological monitoring, remote sensing, and marine science.
Overview
- WildlifeMapper — multi-species detection and identification in aerial survey imagery
- Prairie Dog Detection — aerial detection, mosaicing, and 3D mapping of colonies
- MethaneMapper — hyperspectral methane plume detection
- Underwater & Marine Species — invertebrate detection in deep-sea video and 3D species generation
WildlifeMapper
Detecting, locating, and identifying wildlife from aerial surveys.
WildlifeMapper is a deep learning module for detecting, locating, and identifying multiple animal species in large-scale aerial survey imagery. It replaces the slow, labor-intensive manual counting that underpins wildlife population assessments, giving conservation teams a faster and more consistent way to census animals across vast landscapes.
Unlike earlier detectors that struggle when animals are small and backgrounds are highly uniform, WildlifeMapper introduces a High-Frequency Feature Generator that sharpens the local image structures needed to separate wildlife from terrain. It was trained and validated on a curated dataset of more than 11,000 aerial images and reaches state-of-the-art accuracy across several public aerial survey datasets, improving mean average precision by 42% over prior methods.
- Multi-species detection, localization, and identification from aerial imagery
- High-Frequency Feature Generator for robust detection against homogeneous backgrounds
- Curated and verified dataset of 11k+ aerial images
- State-of-the-art performance across datasets from four countries
- Runs inside BisQue for browser-based analysis
- Built with conservation partners in Kenya and Tanzania
Prairie Dog Detection
Spotting small, rare wildlife — prairie dogs and their burrows — in drone imagery, for grassland conservation.
Prairie dogs are a keystone species: the burrow systems they build and maintain support an entire grassland community, including the endangered black-footed ferret that preys on them. Monitoring prairie dog populations traditionally means manually counting thousands of burrows and hundreds of animals in the field — slow, costly, and error-prone. This module pairs high-resolution drone surveys with AI to automate that count.
The detector is built on RareSpot, a framework combining multi-scale consistency learning with context-aware augmentation to find objects that are tiny (~30 px), sparse, and visually indistinct from the terrain. It improves detection accuracy by more than 35% over baseline models and generalizes to other wildlife datasets. Against expert human annotations it reaches roughly 90% accuracy on burrows; prairie dogs themselves — smaller and more easily hidden — remain harder, at around 50%. A survey that once took days or weeks of fieldwork can now be processed in hours. This work is a collaboration between the UCSB Vision Research Lab and the Smithsonian Institution.
The module runs as a three-stage pipeline that turns a raw drone survey into geolocated detections:
Step 1 — Detection
RareSpot locates individual prairie dogs and burrows across the drone frames, using multi-scale feature alignment to recover objects only ~30 pixels wide against homogeneous grassland.
Step 2 — Mosaicing
Overlapping frames are aligned and stitched into a single georeferenced orthomosaic — one continuous, high-resolution view covering roughly 200 acres.
Step 3 — Mapping
Detections are projected onto the orthomosaic and to geographic coordinates, yielding burrow density, colony extent, and relative abundance that can be tracked over time to guide management.
MethaneMapper
Detecting and quantifying methane plumes in airborne hyperspectral imagery.
Methane is a leading driver of climate change, and much of it leaks from specific, findable sources — pipelines, storage tanks, and oil-and-gas extraction points. Airborne spectrometers such as AVIRIS-NG can map these emissions, but conventional analysis is sensitive to terrain, needs expert inspection, and does not scale. MethaneMapper is an end-to-end spectral-absorption-aware hyperspectral transformer that detects and quantifies methane plumes automatically.
Two novel modules steer the network toward the spectral bands where methane absorbs most strongly, letting it localize plumes accurately across varied land cover. MethaneMapper reaches 0.63 mAP while being five times smaller than the previous state of the art. It is released alongside the Methane HotSpot (MHS) dataset — the largest public hyperspectral methane benchmark, covering 1,200+ ortho-corrected AVIRIS-NG flight lines from 2015–2022 across six states, with over 4,000 annotated plume sites.
Underwater & Marine Species
Detecting invertebrate species in deep-sea ROV video — and generating 3D models of marine fauna.
Each year, underwater remotely operated vehicles (ROVs) record thousands of hours of video from unexplored ocean habitats, but turning that footage into biodiversity data requires scarce, costly expert annotation. This module automates the detection and counting of deep-sea invertebrates directly from ROV video.
It is built on DUSIA (Dataset for Underwater Substrate and Invertebrate Analysis) and a novel context-driven object detector (CDD). CDD uses explicit substrate classification to inform detection — predicting a species class conditioned on the substrate it appears on — which is well suited to the deep sea, where fauna sit against varied seafloor at depths sunlight never reaches. The method also improves training on partially annotated frames and provides a baseline for automated species counting.
Demo
3D species generation
Beyond detection, the module can reconstruct a full 3D model of a species from a single reference frame — for example, generating the fragile pink urchin, with its rigid test and fine spines, from one input image.
DUSIA data download
The DUSIA videos are hosted on the lab's BisQue instance at bisque2.ece.ucsb.edu; request an account and use the Download Manager (export as a TARball) to retrieve the footage.
Contact
BisQue analysis modules are developed in the UCSB Vision Research Lab, led by Professor B.S. Manjunath (Department of Electrical and Computer Engineering, UCSB), in collaboration with the Smithsonian Institution and other partners.