Title:	WABI NBIS Genomics of Seasonal Camouflage in Ptarmigan
DNr:	Berzelius-2026-266
Project Type:	LiU Berzelius
Principal Investigator:	Mafalda Ferreira <mafalda.ferreira@zoologi.su.se>
Affiliation:	Stockholms universitet
Duration:	2026-10-01 – 2027-04-01
Classification:	10610
Keywords:

Abstract

Seasonal camouflage — the moult from brown to white winter plumage — is a key adaptation to snow-covered landscapes that is increasingly mismatched with declining snow cover under climate change. Quantifying how this trait varies across space and time in wild bird populations has been limited by the lack of scalable methods to extract phenotypic measurements from photographs. This project will develop and deploy a deep learning image segmentation pipeline to quantify plumage phenotypes — brightness, colour composition, and moult score — from large, heterogeneous collections of ptarmigan (Lagopus spp.) photographs, including museum specimen images and citizen-science observations from iNaturalist and GBIF. Goals: We will train and validate a segmentation model, to automatically isolate plumage regions and extract quantitative colour and moult metrics from images captured under highly variable photographic conditions (lighting, background, specimen pose, image quality). The resulting phenotypic dataset — spanning over 100 years and several thousand images — will be matched to georeferenced collection localities and climate variables to model the environmental predictors of seasonal colour morph distribution across the Lagopus genus, and will be integrated with population genomic data to identify the genetic architecture underlying loss of camouflage plasticity. Importance. This computer-vision pipeline is being developed as part of a WABI-NBIS collaboration with the BioImage Informatics unit, as the rate-limiting step for a larger project examining the genomic and phenotypic basis of camouflage loss in ptarmigan, funded through a SciLifeLab Fellowship. Without GPU-accelerated training and inference at scale, the phenotyping step cannot proceed, blocking two of the project's four aims (temporal phenological analysis and genotype–phenotype matching). Expected outcomes. By the end of the project period we expect to have: (1) a validated segmentation and phenotyping model, benchmarked against manually scored specimen and citizen-science images; (2) a complete phenotypic dataset covering the ~1,500 planned museum specimens and ~5,000+ citizen-science records; and (3) the public release of the pipeline as an open-source, documented Python package (PtarmiganCV, MIT licence) with a no-code web interface, intended for reuse by any researcher working with specimen or citizen-science photographs of birds, mammals, or reptiles. Software and methods. The specific model architecture and training pipeline will be determined jointly with the NBIS BioImage Informatics team at project start. We anticipate this will involve a CNN-based image segmentation architecture — such as DeepLabv3+ or a comparable encoder-decoder model — trained with a GPU-accelerated deep learning framework (e.g. PyTorch), alongside standard computer vision libraries for preprocessing and augmentation. Training and inference will run on Berzelius' A100/H200 nodes.