Structured Habitat Attribute Extraction from Camera-Trap Images Using a Vision–Language Model: A Proof-of-Concept Study in a Karst Forest

Source abstract

Camera-trap images used in wildlife monitoring contain environmental information that is often underused. This study developed and evaluated a vision–language-model-based workflow for extracting structured visible habitat attributes from camera-trap photographs collected in a karst forest ecosystem in southern China. The workflow combined image preprocessing, schema-constrained prompting, structured JSON generation, programmatic parsing, logical-consistency validation, and automatic re-inference of invalid outputs. Seven attributes were evaluated: terrain, bare-rock visibility, vegetation layers, vegetation density, water visibility, water type, and visible human disturbance. Qwen3-VL-30B-A3B-Instruct was deployed locally, and its outputs were compared with expert-reference annotations for 102 camera-trap images. The evaluation used one model and one prompt/schema configuration and was designed to assess feasibility rather than cross-model or cross-ecosystem generalizability. All images ultimately produced valid structured records after the validation and retry procedure. Field-level exact agreement ranged from 90.20% to 99.02%. For water visibility, exact agreement was 96.08%, balanced accuracy was 97.56%, and macro-F1 was 94.20%; all 20 water-positive images were correctly identified. Water type achieved 95.10% exact agreement, with 19 of the 20 water-positive images matching the expert reference. Five experts separately verified all images using workbooks pre-populated with the reference labels while the model outputs were visible. In this non-blinded verification, Fleiss’ kappa ranged from 0.885 to 1.000 across the six estimable fields (0.974 for vegetation-layer sets treated as nominal categories); visible human disturbance had 100% observed agreement but no category variation, so kappa was not estimable. However, several reference distributions were strongly imbalanced: 96 of 102 images contained visible bare rock, 99 were assigned dense vegetation, and the validation subset contained no disturbance-positive images. The results therefore demonstrate preliminary feasibility under the evaluated conditions rather than broad generalizability. The proposed workflow provides a structured post-acquisition approach for deriving time-aligned visible habitat information from existing camera-trap imagery.