Carbon Footprint Estimation from Indian Food Images

Source abstract

Food production and consumption are global sources of anthropogenic greenhouse gas (GHG) emissions. To encourage sustainable dietary choices for effective climate change mitigation, accurate estimation of the carbon footprint of meals is important. The conventional approaches to carbon footprint estimation, such as dietary surveys and looking up nutrition databases or carbon intensities from online resources followed by manual food logging, can be tedious, erroneous and challenging to scale. In this paper, we propose an artificial intelligence (AI) structure for estimating the carbon footprint of meals using their food images. Specifically, we develop and evaluate a convolutional neural network (CNN) model with YOLOv8m based architecture for food object detection in Indian meals. We trained a YOLOv8m object detection model on a custom dataset containing 31 Indian food categories comprising 7,438 training images and 974 validation images.. For efficient training and inference, we utilized an NVIDIA GeForce RTX 3050 Laptop GPU and trained the model for 50 epochs. With high values of Precision (85.0%), Recall (86.5%), mean average precision at 0.50 IOU (mAP@50) (90.4%) and 0.50 to 0.95 IOU (mAP@50 95) (60.9%), our model demonstrates impressive performance in detecting food items. We linked the detected food items to known carbon emission factors through the Life Cycle Assessment (LCA) method using a carbon mapping layer. This enabled the AI structure to perform the automated estimation of the meal level GHG emissions. Our food recognition structure can be integrated with future enhancements such as portion size estimation using depth estimation and intelligent low-carbon food recommendation engines. By bringing together computer vision with sustainability analytics, we have demonstrated the potential for developing a strong, scalable foundation for estimating the real-time carbon footprint of meals that can support environmentally conscious dietary decision making through mobile and web-based applications.