A scalable deep-learning pipeline plus geospatial analysis — designed to find the assets responsible for pollution at scale and ground every detection in real-world context.
We employ a comprehensive approach to identifying brick kilns using scalable deep learning, and a spatial analysis to understand the scope of brick manufacturing. The workflow combines a machine-learning pipeline for kiln detection and classification with a spatial analysis to assess the environmental and regulatory implications of kiln distribution. We first train the model to identify brick kilns as primary pollution sources, then employ transfer learning to locate and geo-tag other pollutant sources and present them on our visual map.
Satellite imagery trains a classification model distinguishing between images containing kilns and those that do not. Data is provided in HDF5 with separate train/validation files. High-definition imagery is scraped via APIs to augment and label the IGP training set.
A second model further classifies detected kilns into two shape categories: rectangular and ovular. Trained on dedicated labelled data.
After identifying true positives, the pipeline searches for connected components of adjacent kiln tiles, forming signal masses. Centroids are computed and geolocated, then compared against ground-truth kilns to score precision.
Every component of the pipeline — masks, composites, post-processing, spatial joins — is engineered to turn raw imagery into evidence regulators and researchers can use.
Kiln locations are joined with publicly available geospatial datasets (population, health and education facilities, meteorological data) to analyse how kilns relate to regulated entities, environmental conditions, and population exposure.
The final outputs include geo-identified imagery, the kiln-detection model output, and supporting public datasets used in the spatial analysis.
A novel approach to combining deep learning with spatial analysis for environmental detection and regulatory assessment. By keeping methodology and data open, the project enables independent research and collaboration on the impact of brick manufacturing.
The masking function processes CNN outputs into Class Activation Maps (CAMs), highlighting the image regions that drove each prediction. Post-processing normalises masks and runs morphological opening to reduce noise.
Tiles identified as containing kilns are stitched into a composite image with connected-component analysis grouping nearby detections. Geographical coordinates of the top-left corner are derived from the minimum row/column indices.
Median filtering concentrates signal, contrast enhancement sharpens features, Gaussian smoothing reduces noise — improving the clarity and distinctiveness of kiln signals for downstream analysis.
APAD combines deep learning for initial detection with image-processing refinements and visualisation, producing an asset-level database for tracking pollutants in the IGP with high precision and spatial resolution.