Exploratory Analysis Using Deep Learning for Water-Body Segmentation of Peru's High-Mountain Remote Sensing Images.

Area of Science:

• Environmental Science
• Remote Sensing
• Geospatial Analysis

Background:

• High-mountain lakes are vital freshwater resources and indicators of climate change.
• Monitoring these lakes is crucial for understanding environmental dynamics.
• Remote sensing in high-mountain regions faces unique challenges due to terrain and atmosphere.

Purpose of the Study:

• To explore and compare remote sensing techniques for lake monitoring in the Peruvian Andes.
• To evaluate the performance of three deep learning models (DeepWaterMapV2, WatNet, WaterSegDiff) for lake segmentation.
• To assess the suitability of these models for high-mountain lake monitoring.

Main Methods:

• A new dataset of Landsat-8 imagery (2013-2023) was created for the Ancash and Cuzco regions.
• Three deep learning models (DeepWaterMapV2, WatNet, WaterSegDiff) and Normalized Difference Water Index (NDWI) with Otsu thresholding were compared.
• Quantitative metrics (MIoU, PA, F1 Score) and qualitative analysis were used.

Main Results:

• DeepWaterMapV2 and WatNet demonstrated equivalent, adequate lake segmentation performance in challenging conditions.
• WaterSegDiff showed promising qualitative results for lake segmentation.
• WatNet, with lower computational complexity, is identified as a pertinent model for implementation.

Conclusions:

• Deep learning models, particularly WatNet, are effective for high-mountain lake monitoring in the Andes.
• WatNet offers a computationally efficient solution for lake segmentation in these environments.
• Further temporal analysis highlights WatNet's significant behavior in monitoring specific lakes like Singrenacocha.