BMMLDL/BM_Middle_East_Dataset

Data Collection, Preparation & Feature Extraction for AOD Estimation in the Middle East

Data Collection and Preparation

1. Objective
2. Data Collection
3. Preprocessing of AOD Values
4. Dataset Structure
5. Final Combined Dataset
6. Quality Control and Filtering
7. Outcome

---

1. Objective

The objective of this work is to construct a dataset suitable for estimating Aerosol Optical Depth (AOD) over the Middle East using ground-based (AERONET) and satellite-based (Sentinel-2 A2) observations.

---

2. Data Collection

AOD measurements were obtained from AERONET sites across the Middle East. To ensure consistency and quality, only sites satisfying the following criteria were included:

• Availability of Level 2 AOD values.
• Records extending after 2015, since Sentinel-2 L2A products became broadly available post-2015.
• Valid measurements for AOD at 500 nm and AOD at 675 nm wavelengths.

Corresponding Sentinel-2 images were downloaded using a 4 km × 4 km bounding box centered at each AERONET site to ensure that the imagery spatially represents the site measurements.

---

3. Preprocessing of AOD Values

Raw AERONET CSV files varied in format, and most sites did not provide daily averages. Therefore, a preprocessing pipeline was applied as follows:

1. For each site, extract only the following fields:

   • Date
   • Site location (site name / identifier)
   • Daily average of AOD_500nm
   • Daily average of AOD_675nm

2. Compute the daily average AOD at 550 nm using the Ångström interpolation between the 500 nm and 675 nm wavelengths, since 550 nm is the standard reference wavelength for many AOD studies. The Ångström relation was applied per-day to produce a harmonized AOD_550nm value for each observation date.

This procedure produces a single comparable daily AOD value per site-date and eliminates format inconsistency across AERONET CSV files.

---

4. Dataset Structure

For each AERONET site a directory was created with the naming convention:

BM_{sitename}

Each site directory contains:

• train_images/
• test_images/
• BM_{sitename}_train_dataset.csv
• BM_{sitename}_test_dataset.csv

Each CSV follows the column structure:

File,location,aod,path,date,latitude,longitude

Example row (site-level CSV):

train_1.tif,Cairo_EMA_2,0.3821,/Train_images/train_1.tif,2021-08-26,30.080767,31.290067

Notes:

• File is the image filename (relative to the site folder).
• location is the site identifier used in this project.
• aod is the interpolated AOD at 550 nm.
• path is the image path relative to the dataset root or the site folder.
• date is the observation date (YYYY-MM-DD).
• latitude, longitude are the site coordinates.
• For each site 80% of the samples are for training, and 20% are for testing.

---

5. Final Combined Dataset

Two aggregate CSV files were created at the dataset root to allow regional-level training and evaluation:

• BM_Middle_East_train_dataset.csv
• BM_Middle_East_test_dataset.csv

These combined files use the same schema:

File,location,aod,path,date,latitude,longitude

Example aggregated-row:

test_1.tif,Cairo_EMA_2,0.3574,/BM_Cairo_EMA_2/Test_images/test_1.tif,2019-07-23,30.080767,31.290067

Note:

• The final csv files made sure that 80% of each site is for training, and 20% of each site for testing. This way, we make sure that there is no site bias.

---

6. Quality Control and Filtering

• Only Level 2 AERONET records were used to ensure cloud-screened, quality-assured measurements.
• Sites with insufficient post-2015 coverage or missing either 500 nm or 675 nm records were excluded.
• Images were checked to ensure the site fell well within the 4 km × 4 km bounding box; anomalous geolocation mismatches were logged and removed.

---

7. Outcome

The result is a harmonized dataset linking AERONET-derived AOD (interpolated to 550 nm) and co-located Sentinel-2 imagery for 17 sites in the Middle East. The dataset supports both per-site model development and combined regional modeling for AOD estimation.

---

license: mit language: - en pretty_name: g