Abstract: Modern slavery remains a staggering global crisis, with an estimated 12.7 million people trapped in forced labor within the "Brick-Kiln-Belt" of South Asia. This paper presents Tiny-Inception-ResNet-v2, a lightweight deep learning architecture designed to identify and geolocate brick kilns using satellite imagery. By merging Inception blocks for multi-scale feature extraction with residual connections, the model achieves a 94.35% F1-score while maintaining a reduced parameter count of 19.6 million. This research provides a scalable, actionable tool for humanitarian organizations to monitor the region and support the UN’s Sustainable Development Goal (SDG) 8 of ending modern slavery by 2025.
Introduction
The "Brick-Kiln-Belt," spanning Afghanistan, Pakistan, India, Bangladesh, and Nepal, is a focal point for bonded labor. Identifying these kilns is the first step toward intervention, yet traditional surveys are slow, expensive, and often dangerous. Automated detection via satellite imagery offers a promising alternative, but it is complicated by the extreme diversity in kiln morphology.
The Challenge of Kiln Detection
Brick kilns vary significantly in shape, size, and color depending on local technology and soil composition. They can appear circular, oval, or rectangular, often mimicking other land-use features like orchards or housing developments. This variability requires a model capable of multi-scale feature extraction to distinguish kilns from similar-looking structures.
Tiny-Inception-ResNet-v2 Architecture
To address these challenges, the researchers developed Tiny-Inception-ResNet-v2. The architecture leverages two key components:
- Inception Blocks: These allow the network to process features at multiple scales (1x1, 3x3, 5x5 convolutions) simultaneously, which is essential for detecting kilns of varying sizes.
- Residual Connections: These facilitate gradient flow, allowing for deeper training without performance degradation.
The "Tiny" optimization involved a rigorous search for the ideal number of blocks, resulting in a configuration of 10, 3, and 3 blocks across three stages. This reduced the parameter count from 54 million in the standard version to just 19.6 million, making it more efficient for large-scale deployment.
Datasets and Scale
The model was trained on a new, open-sourced dataset of 6,827 manually labeled images across 11 classes (e.g., kilns, roads, farms, parking lots). For testing, the model processed over 787,100 images covering a massive 1.5 million $km^2$ area from Afghanistan to Nepal.
Results and Impact
The Tiny-Inception-ResNet-v2 outperformed significantly larger state-of-the-art networks. Its high recall (90.52%) is particularly valuable for humanitarian monitoring, ensuring that very few kilns are missed during regional sweeps.
| Network Architecture | Precision | Recall | F1-Score | # of Parameters |
|---|---|---|---|---|
| ResNet-152 | 0.9906 | 0.8166 | 0.8952 | 41.4M |
| Inception-v3 | 0.9846 | 0.7413 | 0.8458 | 21.8M |
| Tiny-Inception-ResNet-v2 | 0.9854 | 0.9052 | 0.9435 | 19.6M |
Conclusion
By providing precise geolocation data, this model allows authorities to cross-reference kiln locations with census data on child labor and school attendance. While environmental factors like cloud cover and sensor variation remain challenges, this work demonstrates how optimized deep learning can provide the actionable data necessary to combat modern slavery at scale.
Summary updated