A Multi-Agent AI Workflow for Approximate Carbon Footprint Estimation of Real-World Objects from Images

Sammanfattning

Traditional product carbon footprint assessment relies on detailed lifecycle data that are often unavailable in consumer-facing or early-stage decision contexts. This thesis investigates whether approximate cradle-to-gate carbon footprint estimates can be generated directly from product images using a modular artificial intelligence workflow. A quantitative experimental approach is adopted. The proposed system integrates object detection, vision–language modeling, retrieval-based emission factor access, and modular aggregation to infer product category, material attributes, and corresponding emission factors from image input. The workflow produces approximate embodied carbon footprint estimates together with traceable intermediate outputs. The system is evaluated using a dataset of consumer product images with reference carbon footprint values derived from lifecycle assessment literature. Object detection performance is assessed using precision, recall, and F1-score metrics, while carbon footprint estimation accuracy is evaluated using mean absolute percentage error (MAPE) and absolute error metrics. A category-level baseline model is used for comparison, and statistical robustness is assessed using bootstrapping and confidence interval estimation. Under controlled evaluation conditions, the workflow achieves a MAPE of 7.98%. However, bootstrapped analysis yields a mean MAPE of 39.08% with a 95% confidence interval of [31.69%, 47.34%], indicating sensitivity to dataset variability and evaluation assumptions. The results demonstrate that image-based carbon footprint estimation is feasible, particularly for material-dominated products, while performance remains limited for complex products with non-visible lifecycle factors. The findings highlight a trade-off between representational richness and estimation stability, and emphasize the impact of using category-level reference values when evaluating product-level predictions. The proposed workflow establishes a proof-of-concept for perception-driven environmental impact estimation under conditions of incomplete product metadata.