Usernameavailable283/water-turbidity-random-forest

Water Turbidity Prediction Using Sentinel-2

A machine learning project that uses Sentinel-2 spectral reflectance features and a Random Forest regression model to estimate water turbidity in Formazin Nephelometric Units (FNU).

Overview

This project explores satellite-based water turbidity estimation using Sentinel-2 spectral data and water-quality measurements. It includes dataset cleaning, feature selection, model training, and site-grouped cross-validation.

Highlights

• Model: Random Forest Regressor
• Input: 10 Sentinel-2 spectral reflectance features
• Target: Water turbidity (FNU)
• Cleaned dataset: 233,720 observations
• Sampling sites: 44
• Validation: Five-fold GroupKFold cross-validation by sampling site
• Cross-validation R²: 0.632 on the log1p-transformed target scale
• Model artifact: turbidity_rf_model.pkl

Model Information

Property	Details
Algorithm	Random Forest Regression
Framework	scikit-learn
Number of trees	300
Input features	10 Sentinel-2 spectral features
Prediction target	Turbidity in FNU
Target transformation	log1p(y)
Validation strategy	Five-fold GroupKFold by sampling site

The target transformation reduces the influence of the highly skewed turbidity distribution during model training and evaluation.

Input Features

The model uses the following Sentinel-2 spectral reflectance features:

Feature	Sentinel-2 band
B02_buf250_mean	B02 - Blue
B03_buf250_mean	B03 - Green
B04_buf250_mean	B04 - Red
B05_buf250_mean	B05 - Vegetation red edge
B06_buf250_mean	B06 - Vegetation red edge
B07_buf250_mean	B07 - Vegetation red edge
B08_buf250_mean	B08 - Near-infrared
B8A_buf250_mean	B8A - Narrow near-infrared
B11_buf250_mean	B11 - Short-wave infrared
B12_buf250_mean	B12 - Short-wave infrared

The buf250_mean features represent buffer-based mean reflectance values in the source dataset.

Dataset

The cleaned dataset is derived from water-quality data attributed to Grady P. Ball and Scott D. Ducar (2025), available through the U.S. Geological Survey.

Source: USGS — DOI: 10.5066/P1A7T4FV

Dataset Summary

Metric	Raw file 1	Raw file 2	Cleaned dataset
Rows	874,234	874,233	233,720
Columns	85	85	86
Unique sampling sites	34	26	44
Rows missing at least one selected spectral feature	227,464	167,521	0
Rows missing target values	0	0	0
Exact duplicate rows	0	1	1

Cleaning Results

• No missing values in the ten selected spectral features or the target.
• No negative turbidity values.
• 321 observations with zero turbidity.
• One exact duplicate row in the cleaned dataset.
• 44 unique sampling sites in the cleaned dataset.

Turbidity Distribution

Statistic	Value
Observations	233,720
Mean	20.48 FNU
Median	8.5 FNU
Standard deviation	46.25 FNU
25th percentile	2.5 FNU
75th percentile	22.5 FNU
90th percentile	53.2 FNU
95th percentile	75.8 FNU
99th percentile	159 FNU
Maximum	4,000 FNU

Model Evaluation

The model was evaluated using five-fold GroupKFold cross-validation. Observations were grouped by sampling site, ensuring that validation sites were excluded from the corresponding training folds.

Pooled Out-of-Fold Results

Metric	Result
R² on the log1p-transformed target	0.6321
MAE on the log1p-transformed target	0.5015
Median absolute error on the log1p-transformed target	0.3409
RMSE on the log1p-transformed target	0.7293
Explained variance on the log1p-transformed target	0.6351
MAE in FNU	9.33 FNU
Median absolute error in FNU	2.96 FNU
RMSE in FNU	39.83 FNU

The reported R² is calculated on the log1p-transformed target scale. It is not a percentage-accuracy measurement.

Cross-Validation Results by Fold

Fold	R² on log1p target	MAE (FNU)	RMSE (FNU)
1	0.4322	7.13	32.18
2	0.5714	7.88	16.03
3	0.7545	8.35	17.60
4	0.6946	12.29	26.25
5	0.0908	10.99	75.09

Performance varies across held-out sampling sites.

Baseline Comparison

Model	MAE
Random Forest	9.33 FNU
Training-fold median baseline	17.99 FNU

The Random Forest achieved a lower pooled MAE than the training-fold median baseline.

Performance by Turbidity Range

Actual turbidity range	Observations	MAE (FNU)	RMSE (FNU)
≤ 5 FNU	87,847	3.33	8.37
> 5 to 20 FNU	82,242	5.05	7.72
> 20 to 100 FNU	57,178	13.96	20.73
> 100 FNU	6,453	104.43	227.92

Prediction errors are larger for high-turbidity observations.

Example Usage

The repository includes the trained model artifact turbidity_rf_model.pkl.

The following Google Colab example loads the model and generates one prediction using placeholder feature values.

Install Dependencies

!pip install numpy pandas scikit-learn

Load the Model and Predict

import pickle
import numpy as np
import pandas as pd
from google.colab import files

# Upload turbidity_rf_model.pkl.
uploaded = files.upload()
MODEL_PATH = next(iter(uploaded))

FEATURES = [
    "B02_buf250_mean",
    "B03_buf250_mean",
    "B04_buf250_mean",
    "B05_buf250_mean",
    "B06_buf250_mean",
    "B07_buf250_mean",
    "B08_buf250_mean",
    "B8A_buf250_mean",
    "B11_buf250_mean",
    "B12_buf250_mean",
]

with open(MODEL_PATH, "rb") as f:
    model = pickle.load(f)

# Example input only 
example_input = pd.DataFrame(
    [[0.05] * len(FEATURES)],
    columns=FEATURES,
)

example_prediction = model.predict(example_input)
predicted_turbidity_fnu = np.expm1(example_prediction[0])

print(f"Example predicted turbidity: {predicted_turbidity_fnu:.2f} FNU")

The example assumes the saved model predicts log1p-transformed turbidity and expects features prepared as in the training workflow. The placeholder values are illustrative and do not represent measured Sentinel-2 observations.

Limitations

• Performance varies across held-out sampling sites.
• Prediction errors increase at high turbidity levels.
• Geographic, seasonal, and environmental differences may affect generalization.

Intended Use

• Satellite-based water-quality research.
• Machine learning experimentation.
• Environmental data analysis.
• Educational demonstrations of Random Forest regression.

The model is not a substitute for laboratory testing or independently validated water-quality measurements.

Acknowledgements

The underlying water-quality dataset is attributed to Grady P. Ball and Scott D. Ducar (2025) and is available through the U.S. Geological Survey.

Dataset DOI: 10.5066/P1A7T4FV

The source dataset is identified as CC0 1.0. Original measurements are attributed to their source.

License

This repository uses the MIT License for the project artifacts covered by that license. The original dataset remains subject to its own licensing and attribution terms.