A convolutional neural network that classifies weather photos as Cloudy, Rain, Shine or Sunrise, served through a Streamlit web app.
This was my final-year B.Tech team project (four members), and I did most of the development. After graduating I revisited it: I fixed a modelling bug, evaluated it on a proper held-out test set, and made the app robust to real-world images.
The model was evaluated on a held-out test set of 169 photos that it never saw during training or early stopping.
| Metric | Score |
|---|---|
| Test accuracy | 92.9% |
| Macro F1 | 0.925 |
| Class | Precision | Recall | F1 | Test photos |
|---|---|---|---|---|
| Cloudy | 0.97 | 0.80 | 0.88 | 45 |
| Rain | 0.97 | 0.91 | 0.94 | 33 |
| Shine | 0.83 | 1.00 | 0.90 | 38 |
| Sunrise | 0.96 | 1.00 | 0.98 | 53 |
The most common mistake is Cloudy photos predicted as Shine (8 of 45). Bright, partly cloudy
skies look a lot like sunny ones. All the numbers above come from
reports/metrics.json, which train.py writes.
Training and validation curves stay close together, so the model is not simply memorising the training photos.
- Data: the Multi-class Weather Dataset (CC BY 4.0) has 1,125 photos: Cloudy 300, Rain 215, Shine 253, Sunrise 357. It is split 70/15/15 into training, validation and test sets. The split is stratified, so each set keeps the same class proportions.
- Preprocessing (
preprocess.py): OpenCV decodes each image, converts it to RGB and resizes it to 224×224. Training and the app share this one function, so the model sees identical inputs in both places. - Model (
train.py):- Light augmentation during training only: horizontal flips, small rotations, zoom.
- Rescaling to 0–1 inside the model, so the app can't forget it.
- Four Conv2D + MaxPooling blocks with 32, 64, 128 and 256 filters, then global average pooling, Dense(256), Dropout(0.3), and a softmax output over the 4 classes.
- Adam optimiser and sparse categorical cross-entropy loss. Early stopping on validation loss restores the best epoch.
- App (
app.py): upload a photo and see the predicted condition, its confidence, and the probability for every class.
| Problem in the original notebook | Fix |
|---|---|
The output layer used sigmoid, but the loss used from_logits=True. Those don't agree: the loss treated probabilities as raw scores. |
softmax output with from_logits=False, the correct pairing for mutually exclusive classes |
| Accuracy (89%) was measured only on the validation set, with no separate test set. | A separate held-out test set, plus per-class metrics and a confusion matrix |
Flatten → Dense(256) held 12.8M weights, making a 156 MB model |
Global average pooling: 455K weights, a 5.5 MB model |
| The app crashed on PNGs with transparency, and on greyscale images | cv2.IMREAD_COLOR always yields 3 channels; unreadable files show an error message |
| The app reloaded the model on every interaction | The model is cached with st.cache_resource |
The original notebook is kept for reference in notebooks/.
git clone https://github.com/M-Bhargavi8/ClimateVision.git
cd ClimateVision
python -m venv .venv
.venv\Scripts\activate # macOS/Linux: source .venv/bin/activate
pip install -r requirements.txtUse the trained model: download weather_cnn.keras from the
latest release into models/. Then run:
streamlit run app.pyOr train it yourself: download the dataset from Kaggle and extract it so the photos are in
data/Multi-class Weather Dataset/<Class>/. Then run:
python train.py # about 25 minutes on a laptop CPU; writes models/ and reports/Results can vary slightly between runs and machines, even with the fixed seed.
app.py Streamlit app
train.py training, evaluation and report generation
preprocess.py image decoding shared by training and the app
models/ class_names.json (the .keras model is a release asset)
reports/ metrics.json, confusion_matrix.png, training_curves.png
notebooks/ the original exploration notebook
- 1,125 photos is a small dataset. Expect lower accuracy on photos that look unlike it, such as night scenes, snow or fog. Those aren't among the four classes, but the model will still pick one of them.
- The classes are moderately imbalanced. Rain has the fewest photos (215) and Sunrise the most (357).

