Continuous Satellite Monitoring of Reservoir Capacity Loss Using Deep Learning and Stochastic Mapping: The Poechos Reservoir and Regional Transferability in Northern Peru

DATASET AND SCRIPTS FOR:
Continuous Satellite Monitoring of Reservoir Capacity Loss Using Deep Learning and Stochastic Mapping: The Poechos Reservoir and Regional Transferability in Northern Peru

Authors:
Juan Carlos Breña Aliaga (1,*), Joel Cruz Machacuay (1), Jorge Luis Breña Ore (2), Oscar Felipe (3), Luc Bourrel (4), Pedro Rau (5), and Waldo Lavado-Casimiro (3)

Affiliations:
(1) National Water Authority (ANA), Lima 15036, Peru
(2) Faculty of Chemical and Textile Engineering, National University of Engineering (UNI), Lima 15333, Peru
(3) National Service of Meteorology and Hydrology of Peru (SENAMHI), Lima 15072, Peru
(4) UMR 5563 Géosciences Environnement Toulouse (GET), Université de Toulouse, CNRS, IRD, UPS, CNES, OMP, 14 Avenue Edouard Belin, 31400 Toulouse, France
(5) Centro de Investigación y Tecnología del Agua CITA, Universidad de Ingeniería y Tecnología UTEC, Lima 15063, Peru

Related Publication:
This dataset is associated with the manuscript "Continuous Satellite Monitoring of Reservoir Capacity Loss Using Deep Learning and Stochastic Mapping: The Poechos Reservoir and Regional Transferability in Northern Peru" (currently under peer review).

1. Overview

This repository contains the complete dataset, high-resolution reference masks, deep learning models, and Python scripts required to reproduce the methodology and results presented in the associated research article. The primary framework reconstructs the Elevation-Area-Volume (EAV) curve of the Poechos Reservoir dynamically, resolving the orbital asynchrony between Sentinel-1 SAR imagery and SWOT altimetry using an FPN-InceptionV4 segmentation ensemble and Monte Carlo stochastic quantile mapping.

A supplementary, self-contained module (Folder 7) provides the independent zero-shot regional transferability dataset. This module validates the framework's scalability across three additional Andean-coastal reservoirs (San Lorenzo, Tinajones, and Gallito Ciego) without any local algorithmic fine-tuning.

2. Repository Structure

The repository is organized into 7 compressed folders (.zip):

• 1_Scripts_and_Code.zip — Core Jupyter Notebooks for the Poechos case study.
  • 01_PlanetScope_GroundTruth.ipynb
  • 02_SAR_Preprocessing.ipynb
  • 03_DeepLearning_Segmentation.ipynb
  • 04_Stochastic_Quantile_Mapping.ipynb
• 2_DeepLearning_Dataset.zip — Training/validation data for the primary study area (Poechos).
  • Sentinel_1_Tensors/
  • PlanetScope_Masks/
• 3_Trained_Models.zip — PyTorch state dictionaries (.pth).
  • Best_Ensemble_FPN_InceptionV4/
• 4_SWOT_Altimetry.zip
  • SWOT_Validated_Observations.csv
• 5_InSitu_Validation_Data.zip
  • ANA_Operational_Volumes.csv (historical storage records for Poechos)
• 6_Study_Results_and_Metrics.zip — Tabular data supporting the primary article's figures.
• 7_Supplementary_Regional_Validation.zip — A fully self-contained module for the multi-reservoir zero-shot validation.
  • Scripts/
    • 05_ZeroShot_San_Lorenzo.ipynb
    • 06_ZeroShot_Tinajones.ipynb
    • 07_ZeroShot_Gallito_Ciego.ipynb
  • Datasets/ (Model Inputs) — subfolders San_Lorenzo/, Tinajones/, and Gallito_Ciego/, each including:
    • Sentinel_1_Tensors/: normalized SAR patches (VV, VH, VV−VH) for inference
    • PlanetScope_GroundTruth/: high-resolution 3 m binary water masks
    • SWOT_Altimetry_Filtered.csv: datum-corrected WSE observations
    • ANA_Operational_Records.csv: historical in-situ storage data
  • Transferability_Metrics/ (Generated Outputs) — subfolders San_Lorenzo/, Tinajones/, and Gallito_Ciego/, each including:
    • Sentinel1_Inference_Areas.csv: auto-generated water areas computed by the FPN-InceptionV4 ensemble
    • Spatial_Segmentation_Metrics.csv: pixel-level validation metrics (mIoU, F1-Score)
    • Stochastic_EAV_Curve.csv: final derived capacity curve via quantile mapping
    • Validation_Plots/: multi-panel EAV curves and 1:1 scatter plots (with embedded NSE, RMSE, and BIAS metrics)

3. System Requirements and Dependencies

To execute the scripts locally or on cloud platforms (e.g., Google Colab), the following core libraries are required:

• Python ≥ 3.9
• PyTorch ≥ 2.0
• Segmentation Models PyTorch (segmentation_models.pytorch)
• GeoPandas, Rasterio, NumPy, Pandas, SciPy, scikit-learn

4. Usage Instructions

1. Environment Setup: Install the required dependencies.
2. Primary Study (Poechos): Unzip folders 1 to 6. Update the relative file paths in notebooks 01 to 04 and run them sequentially to replicate the core methodology.
3. Regional Validation: Unzip 7_Supplementary_Regional_Validation.zip. The notebooks 05, 06, and 07 inside the Scripts/ directory can be executed independently. They will automatically read the corresponding inputs from Datasets/ and populate the Transferability_Metrics/ folder with the final tabular data and validation figures.

5. License and Citation

This dataset and code are provided under the Creative Commons Attribution 4.0 International (CC–BY 4.0) license.

If you use these materials in your research, please cite the corresponding peer-reviewed article. (Citation details will be updated upon final publication.)