AUTHOR=Zhang Eddie , Zhang Evan TITLE=Gas pipeline leakage detection based on multiple multimodal deep feature selections and optimized deep forest classifier JOURNAL=Frontiers in Environmental Science VOLUME=Volume 13 - 2025 YEAR=2025 URL=https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2025.1569621 DOI=10.3389/fenvs.2025.1569621 ISSN=2296-665X ABSTRACT=Gas pipeline leaks contribute to one-third of methane emissions annually, posing environmental damage and safety risks. However, accurate and timely detection of the leak presents several challenges, including remote and inaccessible environments, low accuracy and efficiency, and high hardware and labor costs. To address these challenges, we propose a gas pipeline leakage detection architecture based on multiple multimodal deep feature selections and the optimized Improved Deep Forest Classifier (IDFC). First, the multimodal data, thermal images and gas sensor data, are pre-processed. Then a deep feature pool is constructed using the selected Convolutional Neural Network (CNN) models, including AlexNet, ResNet-50, MobileNet, VggNet, and EfficientNet, as well as their inner layers. Aided by the heatmaps created using Gradient-weighted Class Activation Mapping (Grad-CAM), a visualization-based primary feature selection is applied to determine the best features to form an initial CNN pool. The output of the flattened features from this CNN pool is then fed into the IDFC for classification. Hyperparameters of the base learners are then selected for an explainable and enhanced diversity tree-structured deep forest classifier, using the selected multimodal features as inputs. Finally, the Accuracy-Size Comprehensive Indicator (ASCI) is introduced for the secondary feature selection and the optimized deep forest classifier construction, which balances the model accuracy and size and reduces hardware resource requirements. Using the simulated testing dataset created for this research, our architecture demonstrated superior accuracy (98.9%) and deployability with its lower model size (115 MB). This lightweight AI architecture is successfully deployed on a soft robotic system for real-time gas leak detection. Our proposed architecture is also extensible to other environmental hazard detection situations, such as liquid leaks in the pipelines.