hyvioq-v1.1.0

Hyvioq/hyvioq-v1.1.0

hyvioq v1.1.0 is a LightGBM binary classification model that predicts flood severity from precipitation and catchment characteristics, using the INDOFLOODS dataset. It is the second released model checkpoint from hyvioq, an ongoing flood-prediction project that also includes a companion mobile application and the earlier hyvioq-v1.0.0 XGBoost regression model.

The model predicts one of two classes:

• Flood
• Severe Flood

This model was independently developed and trained by Bindupautra Jyotibrat.

Model Details

Model Description

• Developed by: Bindupautra Jyotibrat (BJyotibrat)
• Shared by: BJyotibrat
• Model type: LightGBM Classifier (binary)
• Language(s): Not applicable (tabular model)
• License: GPL-3.0

Model Sources

• Model: BJyotibrat/hyvioq-v1.1.0
• Repository: Jyotibrat/hyvioq
• Evaluation results: Results/LightGBM
• Training dataset: INDOFLOODS

Uses

Direct Use

The model classifies flood events as either Flood or Severe Flood based on precipitation and catchment characteristics.

Downstream Use

The model can be integrated into research and software systems for experimental flood-severity prediction and flood-risk analysis, including as one component of the broader hyvioq mobile app.

Out-of-Scope Use

The model should not be used as a replacement for official flood warnings, professional hydrological analysis, emergency-management systems, or other safety-critical decision-making systems.

Bias, Risks, and Limitations

The model was trained using the INDOFLOODS dataset and may reflect the distribution and characteristics of that dataset.

Model performance may vary for flood events, regions, or catchment characteristics not well represented in the training data.

The model should not be considered a definitive source for real-world flood warnings.

Recommendations

Users should consider the model's evaluation results and the characteristics of the training dataset before using it for research or downstream applications.

For safety-critical applications, predictions should be verified using appropriate hydrological data, expert analysis, and official sources.

How to Get Started with the Model

The model takes an input CSV file of precipitation and catchment characteristics, performs the same feature engineering used during training, and predicts Flood or Severe Flood at a classification threshold of 0.40.

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

# Upload input.csv
print("Upload your input.csv file:")
uploaded = files.upload()

input_file = next(iter(uploaded))

# Load model
model = joblib.load("lightgbm_flood_model.pkl")

# Load user input
data = pd.read_csv(input_file)

# Feature engineering
data["Rain_3day"] = data["T3d"]
data["Rain_5day"] = data["T5d"]
data["Rain_7day"] = data["T7d"]
data["Rain_10day"] = data["T10d"]

data["Rain_1to3_ratio"] = data["T1d"] / (data["T3d"] + 1e-6)
data["Rain_growth_3to7"] = data["T7d"] - data["T3d"]

data["Rain7_Area"] = data["Rain_7day"] * data["Drainage Area"]
data["Rain7_Rugged"] = data["Rain_7day"] * data["Ruggedness Number"]
data["Rain7_Relief"] = data["Rain_7day"] * data["Catchment Relief"]

data["Log_Drainage_Area"] = np.log1p(data["Drainage Area"])

# Features used during training
features = [
    "Rain_3day",
    "Rain_5day",
    "Rain_7day",
    "Rain_10day",
    "Rain_1to3_ratio",
    "Rain_growth_3to7",
    "Drainage Area",
    "Catchment Relief",
    "Drainage Density",
    "Ruggedness Number",
    "Rain7_Area",
    "Rain7_Rugged",
    "Rain7_Relief",
    "Log_Drainage_Area"
]

X = data[features]

# Prediction
probability = model.predict_proba(X)[:, 1]

# Same threshold used during training
prediction = (probability > 0.40).astype(int)

# Display results
for i, (prob, pred) in enumerate(zip(probability, prediction), 1):
    print(f"\n===== Prediction {i} =====")
    print(f"Severe Flood Probability: {prob:.4f}")
    print(
        "Prediction:",
        "Severe Flood" if pred == 1 else "Flood"
    )

Note

Written for Google Colab via google.colab.files.upload(); for local use, replace the upload step with a direct file path, e.g. input_file = "input.csv" and replace lightgbm_flood_model.pkl with flood_model.pkl

Required input columns: T1d, T3d, T5d, T7d, T10d, Drainage Area, Catchment Relief, Drainage Density, Ruggedness Number.

Example Run

An example input.csv used for the run below is available at Inferencing/input.csv.

Training Details

Training Data

The model was trained using the INDOFLOODS dataset, created by Sai Kiran Kuntla and Manabendra Saharia.

The original INDOFLOODS dataset is licensed under CC BY 4.0. Users of the underlying dataset should follow its original license and attribution requirements.

Training Procedure

The model was trained using LightGBM with engineered precipitation and catchment features.

Preprocessing

The preprocessing pipeline included:

• Rainfall aggregation features
• Rainfall ratio and trend features
• Rainfall–catchment interaction features
• Log transformation of drainage area
• Removal of unused raw rainfall features
• Removal of latitude and longitude from the model features

Training Hyperparameters

• Model: LGBMClassifier
• Number of estimators: 800
• Maximum depth: 3
• Learning rate: 0.03
• Subsample: 0.8
• Column sampling by tree: 0.8
• Class weighting: scale_pos_weight
• Random state: 42
• Classification threshold: 0.40
• Training regime: Standard precision

Speeds, Sizes, Times

Training and evaluation were performed on Google Colab in under one hour.

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was evaluated on a test split created from the INDOFLOODS dataset, using an 80/20 train-test split with stratification and random_state=42. A 5-fold stratified cross-validation procedure was also performed.

Factors

The evaluation considers binary flood-severity classification between Flood and Severe Flood.

Metrics

Accuracy, precision, recall, F1 score, ROC AUC, confusion matrix, classification report, 5-fold cross-validation, and feature importance.

Results

ROC AUC Curve:

Confusion Matrix:

Classification Report:

              precision    recall  f1-score   support

           0       0.85      0.63      0.72       584
           1       0.55      0.81      0.65       326

    accuracy                           0.69       910
   macro avg       0.70      0.72      0.69       910
weighted avg       0.74      0.69      0.70       910

Full evaluation logs and results are available at Results/LightGBM.

Summary

The model achieves a ROC AUC of 0.80 and recall of 0.81 on the Severe Flood class, indicating it catches most severe-flood events in the test split, at the cost of lower precision (0.55) — i.e. it over-predicts severity more often than it misses it. These results are from the held-out evaluation split and should not be interpreted as guaranteed real-world performance.

Model Examination

Feature importance analysis was performed to examine the relative importance of the features used by the LightGBM model. Detailed results are available at Results/LightGBM.

Environmental Impact

Carbon emissions can be estimated with the ML Impact calculator (Lacoste et al., 2019).

• Hardware Type: NVIDIA T4 GPU
• Hours used: < 1 hour
• Cloud Provider: Google Colab

Technical Specifications

Model Architecture and Objective

LightGBM gradient-boosted decision-tree classifier. The objective is binary classification of flood severity into Flood and Severe Flood.

Compute Infrastructure

• Hardware: NVIDIA T4 GPU
• Software: Python, LightGBM, pandas, NumPy, scikit-learn, joblib

Citation

BibTeX:

@software{jyotibrat_hyvioq_v1_1,
  author  = {Bindupautra Jyotibrat},
  title   = {hyvioq-v1.1.0: LightGBM Flood Severity Classification Model},
  year    = {2026},
  publisher = {Hugging Face},
  url     = {https://huggingface.co/BJyotibrat/hyvioq-v1.1.0}
}

APA:

Jyotibrat, B. (2026). hyvioq-v1.1.0: LightGBM Flood Severity Classification Model. Hugging Face. https://huggingface.co/BJyotibrat/hyvioq-v1.1.0

When using the training data, please also cite the original INDOFLOODS dataset and associated research.

Glossary

• LightGBM — a gradient boosting framework based on decision-tree algorithms.
• Flood — the lower-severity class predicted by the model.
• Severe Flood — the higher-severity class predicted by the model.
• ROC AUC — area under the Receiver Operating Characteristic curve.
• F1 Score — harmonic mean of precision and recall.
• INDOFLOODS — a flood-event dataset containing flood information and catchment characteristics for India.

More Information

• Evaluation results: Results/LightGBM
• Training dataset: INDOFLOODS

Authors

• Bindupautra Jyotibrat (BJyotibrat)

Contact

Email: [email protected]