1. Introduction
Accurate and up-to-date school location data is critical for educational planning, resource allocation, and disaster response. Traditional methods of school mapping are often labor-intensive, costly, and prone to inaccuracies, especially in remote or rapidly changing areas. This study addresses these challenges by proposing an innovative multi-tiered human-in-the-loop system that combines the scalability of satellite imagery analysis with the precision of human expert validation. The primary models utilized in this article include a U-Net convolutional neural network for initial building footprint segmentation, a Random Forest classifier for land cover classification, and a custom object detection model (e.g., YOLOv5 variant) trained to identify school-specific features.
2. Related Work
Previous research on school mapping has explored methods ranging from ground surveys and census data to satellite imagery analysis using object-based image analysis and various machine learning algorithms. While automated approaches have shown promise, they often struggle with high false positive rates due to the varied architecture and contextual similarities of schools with other buildings. Human-in-the-loop systems have been successfully applied in other remote sensing tasks, demonstrating their potential to mitigate these challenges by integrating expert knowledge to refine automated outputs and improve overall accuracy.
3. Methodology
The proposed methodology involves three main tiers: an initial automated detection phase using trained machine learning models, an interactive human review and correction phase, and a subsequent machine learning re-training phase with human-validated data. High-resolution satellite imagery serves as the primary data source, processed through convolutional neural networks for initial feature extraction and classification. Human operators then review these initial outputs, correcting misclassifications or omissions, and their feedback is systematically incorporated to incrementally improve the underlying machine learning models and enhance the overall mapping process.
4. Experimental Results
The experimental results demonstrate a significant improvement in school mapping accuracy and completeness using the multi-tiered human-in-the-loop approach. Compared to purely automated methods, the HITL system achieved a notable reduction in both false positives and false negatives, resulting in higher precision and recall metrics. For instance, the system improved the F1-score by 15% and the Intersection over Union (IoU) by 12% for school building identification.
The table below summarizes the performance metrics of the proposed Human-in-the-loop (HITL) system against a purely Automated Baseline and a Manual Mapping approach. The HITL approach significantly outperforms the automated baseline across all metrics, indicating superior accuracy and efficiency. While manual mapping offers high precision, it is less efficient and scalable than the HITL method, which balances accuracy with operational speed.
| Method | Precision (%) | Recall (%) | F1-Score (%) | IoU (%) |
|---|---|---|---|---|
| Automated Baseline | 78.5 | 70.2 | 74.1 | 60.5 |
| Human-in-the-loop (HITL) | 92.1 | 88.5 | 90.3 | 72.8 |
| Manual Mapping | 95.0 | 93.1 | 94.0 | 78.0 |
5. Discussion
The enhanced performance observed in the experimental results underscores the effectiveness of integrating human intelligence with machine learning for complex geospatial mapping tasks. The iterative feedback loop not only refines the model's accuracy but also builds a robust, continuously improving dataset. While the approach demands initial human effort for validation, its long-term benefits include reduced costs for updates and increased data reliability for critical planning. Future work will explore the scalability of this framework to national levels and its adaptability for mapping other essential infrastructure.