Deep Learning Framework for Large-Scale Precision Mapping of Brick Kilns to Support Sustainable Policy in the Indo-Gangetic Plain Yamini Agrawal1,2,*, Hina Pande1, Poonam Seth Tiwari1, Shefali Agrawal1, Prakash Chauhan3
1 Indian Institute of Remote Sensing,
Indian Space Research Organisation,
Dehradun, Uttarakhand, India
2 Department of Civil Engineering,
Indian Institute of Technology Roorkee,
Roorkee, Uttarakhand, India
3 National Remote Sensing Centre,
Indian Space Research Organisation,
Hyderabad, Telangana, India
Abstract
The urban population in South Asia is projected to grow by 250 million by 2030, driving an expansion of built-up areas and escalating demand for raw construction materials, particularly fired clay bricks. Currently, the region produces approximately 310 billion bricks annually, contributing substantially to particulate matter emissions. However, the absence of a documented, spatially explicit survey of active brick kilns limits policymakers ability to assess and mitigate the sectors environmental impacts. This study applies a novel deep learning-based detection-segmentation approach to map brick kilns in the entire northern Indo-Gangetic Plains (IGP), which contains a high density of kilns on alluvial, sandy, clayey, and loamy soils. A deep learning model was trained, fine-tuned using hyperparameter tuning and tested on multi-sensor optical satellite imagery to detect and segment kiln structures. On the validation set, the model achieved mAP >= 0.50, mask mAP >= 0.87, with inference throughput exceeding 15 frames s-1, demonstrating suitability for large-scale, near real-time applications. Final results yielded an average precision, recall, and F1-score of 0.881, 0.827, and 0.853, respectively, identifying kilns in the study area. Post segmentation, Normalized Difference Vegetation Index (NDVI) and Short-Wave Infrared (SWIR) spectral indices were derived from multi-temporal satellite imagery to track interseasonal and annual dynamical changes, enabling discrimination between active and abandoned brick kilns. Segmentation output provided accurate boundary delineations, enabling estimation of kiln footprint area, potential brick production capacity, and associated carbon footprint. The findings highlight the potential of advanced object detection-segmentation frameworks for automated and scalable environmental monitoring and generating comprehensive spatial inventories. This approach can directly inform regulatory compliance, and support emission-reduction strategies.
Keywords: Brickkiln, deep learning, detection, environment, urban
