# Land Cover Change Detection via Semantic Segmentation

**Authors:** Renee Su, Rong Chen  
**arXiv:** [1911.12903](https://arxiv.org/abs/1911.12903)  
**Date:** 2019-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/1911.12903v1/

## Key points

- Detailed method for detecting land cover changes from aerial imagery using semantic segmentation
- Utilizes Deeplab v3+ as state-of-the-art semantic segmentation technology
- Land cover classification training pipeline includes data preparation, model training, and model exporting modules
- Inputs are images from Google Earth at the same location but from different times
- Trained model predicts semantic segmentation results to identify changes in land cover
- System calculates percentage of each land cover class for every input image
- Improved accuracy compared to previous work, with mean IoU score of 0.756
- Applications in deforestation analysis, land management, and urban planning
- Valuable insights for decision-making related to environmental conservation and urban development

## Summary

This paper presents a detailed method for detecting land cover changes from aerial imagery using semantic segmentation, a machine learning approach. The authors propose a land cover classification training pipeline that utilizes Deeplab v3+, which is state-of-the-art semantic segmentation technology. The pipeline includes data preparation, model training for seven different land cover types, and model exporting modules. In the land cover change detection system, the inputs are images retrieved from Google Earth at the same location but from different times. The trained model is then used to predict semantic segmentation results on these images, allowing for the identification of changes in land cover. Additionally, the system calculates the percentage of each land cover class for every input image. The authors report an improvement in the accuracy of their land cover semantic segmentation model compared to previous work. They achieve a mean Intersection over Union (IoU) score of 0.756, whereas the DeepGlobe land cover classification challenge reported an IoU score of 0.433. This improvement demonstrates the effectiveness of leveraging state-of-the-art semantic segmentation technology for accurate land cover change detection. The proposed method has various applications in fields such as deforestation analysis, land management, and urban planning. By accurately identifying and monitoring changes in land cover over time, this method can provide valuable insights for decision-making processes related to environmental conservation and urban development. Overall, this paper presents a comprehensive approach to detecting land cover changes using semantic segmentation technology. The improved accuracy achieved by this method makes it a valuable tool for analyzing and managing landscapes in various domains.