Methodology
We employ a comprehensive approach to identifying brick kilns using scalable deep learning techniques and conducting a spatial analysis to understand the scale and scope of brick manufacturing. The workflow is organized into two main components: a machine learning pipeline for kiln detection and classification, and a spatial analysis to assess the environmental and regulatory implications of brick kiln distribution. We first train our model to identify brick kilns as primary sources of pollutants. We then employ transfer learning to locate and geo-tag other pollutant sources and present them on our visual map.
Machine Learning Pipeline
Kiln Detection Model Training
After identifying true positives and classifying the images, the pipeline searches for connected components of adjacent kiln images to form signal masses. The centroids of these masses are calculated. This step is designed to extract coordinates of the midpoints of the identified brick kilns. These can then be geolocated and mapped to then compare the model’s prediction accuracy against the existing kilns.
Kiln Detection Model Training
The model utilizes satellite imagery to train a classification model distinguishing between images containing kilns and those that do not. The data is provided in HDF5 format, with separate files for training and validation. The images are first stripped of their geospatial metadata, focusing solely on the visual content (RGB data). For the IGP, high definition satellite imagery is scraped using APIs to augment and label the training dataset.
Shape Classification Model Training
We further classify kilns into two types based on their shape: rectangular / ovular. This step involves training another model with data labeled for these two categories.
Spatial Analysis
This component utilizes the kiln locations identified by the machine learning pipeline along with various publicly available geospatial datasets (e.g., population distributions, health and education facilities, meteorological data) to analyze the impact of brick kilns. The analysis aims to understand the distribution of kilns in relation to regulated entities, environmental conditions, and population exposure to pollution.
Expected Results
The final data will include geo-identified imagery, final output of the kiln detection model, and public datasets for the spatial analysis.
Functional Framework
This framework offers a novel approach to leveraging deep learning and spatial analysis for environmental detecting and regulatory assessment. By making the methodology and data accessible, the project encourages further research and collaboration in addressing the challenges posed by brick manufacturing and its environmental impact in the IGP. For starters, we create a model to detect Brick kilns in the Indian Gangetic Plain. It operates on the assumption that satellite images of a specific region have been processed through a convolutional neural network (CNN) model, which outputs predictions indicating the presence of kilns. The code consists of several key components designed to enhance the model's output, create visualizations, and refine detections.
Mask Images Module to segregate and classify kilns
The masking function processes the outputs of a CNN model to generate Class Activation Maps (CAMs) for each image. These maps highlight the regions within the images that most contributed to the model's predictions, effectively acting as a "mask" that can be used to locate kilns.
The function iterates over the model's predictions and uses the CAM technique to up sample and weight the last convolutional layer's output, creating a heatmap that emphasizes predicted kiln locations.
Post-processing steps, including normalization and morphological operations (like opening to reduce noise), refine the masks, making the signals (kiln locations) clearer.
Create Connected Component Image and geo-tag kilns:
This function will generate a composite image from individual tiles that have been identified as containing kilns. It uses a connected component analysis to group nearby detections, considering them part of the same geographical feature.
The composite image is constructed by pasting the original satellite image tiles and their corresponding masks onto a new image canvas. This process involves calculating the geographical coordinates for the top-left corner of the composite image based on the minimum row and column indices among the detected components.
Post Process Connected Component Mask to improve visibility and robustness
This process further refines the mask of the composite image by applying a series of filters and enhancements. These include median filtering to concentrate signals, contrast enhancement to make features more distinct, and Gaussian smoothing to reduce noise.
The aim is to improve the clarity and distinctiveness of kiln signals in the mask, facilitating easier identification and analysis.
Overall Workflow and Application Summary
APAD demonstrates a sophisticated approach to environmental detecting or industrial activity surveillance, leveraging machine learning to interpret satellite images and produce actionable insights with high precision and spatial resolution. The combination of deep learning for initial detection and sophisticated image processing techniques for refinement and visualization will allow us to create a comprehensive asset level database for tracking pollutants in the IGP.