Near Real-Time Mapping of Tropical Forest Disturbance Using SAR and Semantic Segmentation in Google Earth Engine
, Ate Poortinga 2,3, Biplov Bhandari 4,5, Nyein Soe Thwal 2, Nguyen Hanh Quyen 2, Jeff Silverman 5, Karis Tenneson 4,5, David Bell 6, Matthew Gregory 7, Robert Kennedy 1 and David Saah 2,8Round 1
Reviewer 1 Report
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Reviewer 2 Report
The article presents an intriguing approach on how to monitor forest cover changes using SAR (Synthetic Aperture Radar) technology and deep learning models. It underscores the challenge of persistent cloud cover in continuous monitoring, particularly in tropical regions, and provides a solution to this issue. The article demonstrates the implementation of this method using tools like Google Earth Engine and shows promising results in terms of performance. It also highlights the advantages of this approach, including greater flexibility and ease of development, which are valuable for practical applications from an operational perspective.
1. Limited Performance Comparison: The article lacks detailed performance comparison data with existing systems. To address this, we should conduct more comprehensive performance comparisons to highlight our method's strengths and weaknesses.
2. Overfitting in Deep Learning: The article acknowledges a performance gap between our deep learning test set and photo interpretation dataset, likely due to overfitting. To mitigate this, we need to improve model generalization through strategies like reducing complexity, increasing data augmentation, and applying appropriate regularization techniques.
3. No Spatial Partitioning Strategy: The absence of a spatial partitioning strategy may have concealed overfitting. To rectify this, we should incorporate spatial partitioning to thoroughly assess model performance and identify overfitting issues.
4. Untapped Transfer Learning: Not utilizing transfer learning may limit model performance. To address this, we should explore transfer learning by pretraining on high-quality data from another region and fine-tuning on the target dataset.
5. Missing Continuous Change Detection Algorithms: The article did not explore continuous change detection algorithms to enhance training data quality, especially regarding timing of forest cover changes. We should investigate and implement such algorithms to improve data quality.
By implementing these modifications, we can overcome the limitations, enhancing the research's scientific quality and practicality while opening doors for future improvements.
Author Response
Please see the attachment..
Author Response File: Author Response.pdf
