Air pollution is a significant environmental threat faced by the world today. With each passing day, the air we breathe becomes increasingly contaminated, leading to severe health issues for individuals. Unfortunately, the existing air pollution monitoring stations are both expensive and insufficiently distributed. Consequently, the effectiveness of air pollution mitigation measures is hindered by inadequate data, underscoring the necessity for cost-effective alternatives. This paper introduces a sparse autoencoder-based ensemble model for estimating particulate matter concentrations using outdoor images. First, an L1 regularized sparse convolutional autoencoder compresses and extracts the pertinent features from images. Subsequently, these extracted image features are combined with weather and traffic data, followed by dimensionality reduction through principal component analysis. The final step involves a stacked ensemble of regression models to estimate PM2.5 concentrations. The ensemble incorporates support vector regression, k-nearest neighbor, and random forest regressor as base learners, with the light gradient boost machine acting as the meta-learner. An extensive dataset of 8488 single-scene outdoor images, named 'AirSetTvm,' has been meticulously collected and labeled with corresponding ground truth values derived from the continuous monitoring station. Encouraging results from this model, in comparison with other deep learning models in the existing literature, suggest that the proposed approach is a viable, cost-effective alternative for estimating PM2.5 concentrations.