...

How to build a Keras CNN for Weather Image Classification

/ TensorFlow tutorials
tensorflow image classification

Introduction

This tutorial shows a complete TensorFlow image classification pipeline using Keras CNN.
You will learn how to split a weather photo dataset into train and validation sets.
You will apply image data augmentation with Keras ImageDataGenerator to improve generalization.
You will build and train a deep convolutional neural network for five weather classes.
You will visualize accuracy and loss, save the best model, and run fast batch predictions on test images.

The link for the video tutorial is here : https://youtu.be/gFiISJPCpKs&list=UULFTiWJJhaH6BviSWKLJUM9sg

You can find the full code here : https://ko-fi.com/s/efaafe52c5

More tutorials in my blog : https://eranfeit.net/blog/

Here is the code :

Installation :

 # Requirements : Nvidia GPU card & and Cuda tool kit install # I am using this card : https://amzn.to/3mTa7HX # Working Anaconda enviroment  #dataset :  https://www.kaggle.com/datasets/vijaygiitk/multiclass-weather-dataset  conda create -n weather-predict-CNN python=3.7 conda activate weather-predict-CNN  pip install tensorflow pip install tensorflow-gpu pip install pillow pip install SciPy pip install matplotlib pip install pandas pip install numpy

Link for the dataset :https://www.kaggle.com/datasets/vijaygiitk/multiclass-weather-dataset

Preparing and Splitting the Weather Dataset

This section prepares the directory structure and splits images into training and validation sets.
It ensures only non-empty files are copied and preserves a randomized distribution across the splits.

A reliable train–validation split is the backbone of any TensorFlow image classification workflow.
By scanning each class directory and excluding zero-byte files, the pipeline guards against corrupt samples that could destabilize training or inflate error metrics.
Randomized shuffling ensures that seasonal biases, lighting conditions, or camera angles are evenly distributed across splits, which directly improves generalization for weather image classification.

Using an 85% training and 15% validation ratio provides ample data to learn while preserving a clean hold-out set for monitoring overfitting.
Because the split is performed per class folder, class balance is naturally preserved as long as the source folders are balanced.
If your dataset is imbalanced, consider adjusting the ratio or augmenting underrepresented classes to maintain fair representation during keras cnn training.

The file operations are implemented with os and shutil for speed and simplicity.
Each valid filename is collected, shuffled, and copied into Train and validate destinations that mirror the class taxonomy.
This keeps directory semantics consistent with Keras flow_from_directory, minimizing friction between dataset curation and model ingestion.

Reproducibility is critical in deep learning pipelines.
If you want exact reproducibility, seed Python’s random before sampling and record the seed alongside your experiment notes.
This allows you to regenerate the exact same split and directly compare different augmentation or optimizer settings in later experiments.

Well-structured folders also streamline MLOps tasks such as dataset versioning and storage.
Placing Train and validate under a single dataset root simplifies backup, cloud syncing, and collaboration.
It also makes it trivial to add a dedicated Test folder later for unbiased, final evaluation.

### Import the os module to interact with the filesystem. import os ### Import random for shuffling the dataset deterministically if seeded or randomly for diversity. import random ### Import shutil to copy files between folders. import shutil  ### Define the original dataset folder path that contains class subfolders. dataOrgFolder = "C:/Python-cannot-upload-to-GitHub/Weather/original-dataset/" ### Define the base dataset folder path that will host Train and validate splits. dataBaseFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset"  ### List the class directories inside the original dataset folder. dataDirList = os.listdir(dataOrgFolder) ### Print the list of class directories for verification. print(dataDirList)  ### Set the fraction of data that goes to the training split. splitSize = .85  ### Define a function to split files from a source class folder into training and validation targets. def split_data (SOURCE , TRAINING , VALIDATION , SPLIT_SIZE):     ### Initialize an empty list to collect valid filenames.     files = []      ### Iterate over each filename inside the source directory.     for filename in os.listdir(SOURCE) :         ### Build the absolute path to the file.         file = SOURCE + filename         ### Print the file path for debugging and traceability.         print(file)         ### Check that the file has nonzero size to avoid copying corrupt or empty files.         if os.path.getsize(file) > 0 :             ### Append the valid filename into our list.             files.append(filename)         else:             ### Warn that the file will be skipped due to zero length.             print(filename + " has 0 length , will not copy this file !!")      ### Print the total number of valid files found in the source class folder.     print(len(files))      ### Compute the number of training samples based on the split ratio.     trainLength = int(len(files) * SPLIT_SIZE )     ### Compute the number of validation samples as the remainder.     validLength = int( len(files) - trainLength )      ### Randomly shuffle the dataset to ensure a fair distribution.     suffleDataSet = random.sample(files, len(files))      ### Slice out the training subset from the shuffled list.     trainingSet = suffleDataSet[0:trainLength]     ### Slice out the validation subset from the shuffled list.     validSet = suffleDataSet[trainLength:]      ### Copy each training file from the source to the training destination.     for filename in trainingSet:         ### Build the absolute path to the source file.         f = SOURCE + filename         ### Build the absolute path to the training destination file.         dest = TRAINING + filename         ### Perform the actual file copy.         shutil.copy(f, dest)           ### Copy each validation file from the source to the validation destination.     for filename in validSet:         ### Build the absolute path to the source file.         f = SOURCE + filename         ### Build the absolute path to the validation destination file.         dest = VALIDATION + filename         ### Perform the actual file copy.         shutil.copy(f, dest)   ### Define source, train, and validation folder paths for the 'cloudy' class. cloudySourceFolder = "C:/Python-cannot-upload-to-GitHub/Weather/original-dataset/cloudy/" cloudyTrainFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/Train/cloudy/" cloudyValidFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/validate/cloudy/"  ### Define folder paths for the 'foggy' class. foggySourceFolder = "C:/Python-cannot-upload-to-GitHub/Weather/original-dataset/foggy/" foggyTrainFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/Train/foggy/" foggyValidFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/validate/foggy/"  ### Define folder paths for the 'rainy' class. rainyySourceFolder = "C:/Python-cannot-upload-to-GitHub/Weather/original-dataset/rainy/" rainyTrainFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/Train/rainy/" rainyValidFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/validate/rainy/"  ### Define folder paths for the 'shine' class. shineSourceFolder = "C:/Python-cannot-upload-to-GitHub/Weather/original-dataset/shine/" shineTrainFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/Train/shine/" shineValidFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/validate/shine/"  ### Define folder paths for the 'sunrise' class. sunriseSourceFolder = "C:/Python-cannot-upload-to-GitHub/Weather/original-dataset/sunrise/" sunriseTrainFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/Train/sunrise/" sunriseValidFolder = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/validate/sunrise/"  ### Execute the split for the 'cloudy' class with the selected ratio. split_data(cloudySourceFolder , cloudyTrainFolder , cloudyValidFolder , splitSize) ### Execute the split for the 'foggy' class with the selected ratio. split_data(foggySourceFolder , foggyTrainFolder , foggyValidFolder , splitSize) ### Execute the split for the 'rainy' class with the selected ratio. split_data(rainyySourceFolder , rainyTrainFolder , rainyValidFolder , splitSize) ### Execute the split for the 'shine' class with the selected ratio. split_data(shineSourceFolder , shineTrainFolder , shineValidFolder , splitSize) ### Execute the split for the 'sunrise' class with the selected ratio. split_data(sunriseSourceFolder , sunriseTrainFolder , sunriseValidFolder , splitSize)

You can find the full code here : https://ko-fi.com/s/efaafe52c5

Building and Training a TensorFlow CNN with Keras

This section constructs a multi-block convolutional network with ReLU activations and MaxPooling.
It augments training images, compiles the model with categorical crossentropy, and trains with validation tracking.

This keras cnn architecture stacks multiple Conv2D → ReLU → MaxPooling stages, progressively increasing the number of filters from 16 up to 256.
Early layers learn local features such as edges and color gradients, while deeper blocks capture higher-order structures like textures and silhouettes characteristic of cloudy, foggy, rainy, shine, and sunrise scenes.
The pairings of consecutive convolutions before pooling allow the network to refine features at a given spatial resolution before reducing dimensionality.

Image data augmentation is applied via ImageDataGenerator to boost robustness.
Rescaling to [0, 1] normalizes pixel intensities, while rotation and zoom simulate real-world camera variance that your model will encounter in deployment.
Horizontal flips add viewpoint diversity for landscape scenes, which is especially helpful in weather image classification where left–right symmetry does not alter semantics.

The model is compiled with Adam and categorical_crossentropy, a standard setup for multi-class TensorFlow image classification.
Tracking accuracy alongside the validation metric provides an intuitive signal of learning progress.
If validation accuracy stalls or begins to diverge from training accuracy, it is a sign of overfitting or insufficient augmentation.

ModelCheckpoint monitors val_accuracy and persists the best-performing weights to disk.
This guarantees that even if later epochs degrade due to noise or an overly long schedule, you still keep the strongest version for inference.
Pairing this with EarlyStopping (optionally on val_loss) is a powerful pattern to reduce training time and improve generalization in keras cnn projects.

The deep Dense(512) → Dense(512) head aggregates spatial features learned by the convolutional stack.
These layers model complex class boundaries that single linear projections might miss, especially when weather cues are subtle or partially occluded.
A final softmax outputs calibrated probabilities across the five classes, enabling thresholding, top-k inspection, and clean confusion-matrix analysis.

Training curves for accuracy and loss are plotted with Matplotlib to visualize learning dynamics.
Healthy curves show training and validation accuracy rising together while losses fall, with only modest gaps between them.
Large gaps suggest the need to tune augmentation strength, adjust model depth, or incorporate regularization techniques such as Dropout or weight decay.

### Import Keras utilities for reading images and generating augmented batches. from tensorflow.keras.preprocessing.image import ImageDataGenerator ### Import the main convolutional, pooling, and dense layers. from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout ### Import the Sequential API to stack layers linearly. from tensorflow.keras.models import Sequential ### Import the Adam optimizer for efficient gradient updates. from tensorflow.keras.optimizers import Adam ### Import callbacks to control training and save model checkpoints. from tensorflow.keras.callbacks import EarlyStopping , ModelCheckpoint ### Import matplotlib to visualize training metrics. import matplotlib.pyplot as plt   ### Set the target input width for the model. imgWidth = 256 ### Set the target input height for the model. imgHeight=256 ### Define how many samples per batch to load. batchSize = 16  ### Set the path to the training directory containing class subfolders. TRAIN_DIR = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/Train/" #dont forget the last /  ### Create a training ImageDataGenerator with rescaling and augmentation. train_datagetn = ImageDataGenerator (rescale = 1/255.0,                                     rotation_range = 30 ,                                     zoom_range = 0.4,                                     horizontal_flip=True)  ### Build the training dataset pipeline from the directory. train_data_set = train_datagetn.flow_from_directory(TRAIN_DIR,                                                     batch_size = batchSize,                                                     class_mode = 'categorical',                                                     target_size = (imgHeight,imgWidth))  ### Set the path to the validation directory. VALIDATION_DIR = "C:/Python-cannot-upload-to-GitHub/Weather/dataset/validate/" ### Create a validation ImageDataGenerator with rescaling only. val_datagen = ImageDataGenerator(rescale = 1/255.0) ### Build the validation dataset pipeline from the directory. val_data_set = val_datagen.flow_from_directory(VALIDATION_DIR,                                                     batch_size = batchSize,                                                     class_mode = 'categorical',                                                     target_size = (imgHeight,imgWidth))   ### Instantiate EarlyStopping to halt training if validation loss stops improving. callback = EarlyStopping(monitor='val_loss', patience=5, verbose=1, mode='auto')  ### Define a path to store the best model weights discovered during training. best_model_file_path = "C:/Python-cannot-upload-to-GitHub/Weather/bestWeatherModel.h5"  ### Configure ModelCheckpoint to save the model with the highest validation accuracy. best_model = ModelCheckpoint(best_model_file_path, monitor="val_accuracy", verbose=1, save_best_only=True)  ### Create a Sequential CNN architecture with stacked Conv2D and MaxPooling layers. model = Sequential([     ### First convolution extracts low-level features with 16 filters.     Conv2D(16, (3,3) , activation='relu', input_shape=(imgHeight, imgWidth,3)),     ### Downsample spatial dimensions while retaining salient features.     MaxPooling2D(2,2),     ### Second convolution increases capacity with 32 filters.     Conv2D(32, (3,3) , activation='relu'),     ### Pooling reduces feature map size and provides translation invariance.     MaxPooling2D(2,2),     ### Third convolution deepens feature extraction with 64 filters.     Conv2D(64, (3,3) , activation='relu'),     ### A second 64-filter convolution refines learned patterns.     Conv2D(64, (3,3) , activation='relu'),     ### Pooling halves spatial dimensions again.     MaxPooling2D(2,2),     ### Fourth convolution learns more complex structures with 128 filters.     Conv2D(128, (3,3) , activation='relu'),     ### Another 128-filter convolution to compound features.     Conv2D(128, (3,3) , activation='relu'),     ### Pooling controls overfitting and reduces compute.     MaxPooling2D(2,2),     ### Fifth convolution block with 256 filters captures high-level patterns.     Conv2D(256, (3,3) , activation='relu'),     ### Additional 256-filter convolution to enrich representations.     Conv2D(256, (3,3) , activation='relu'),     ### Third 256-filter convolution for deeper abstractions.     Conv2D(256, (3,3) , activation='relu'),     ### Final pooling before flattening to feed dense layers.     MaxPooling2D(2,2),     ### Flatten converts 3D feature maps into a 1D vector.     Flatten(),     ### Dense layer with 512 units for non-linear combination of features.     Dense(512, activation='relu'),         ### Another Dense(512) to enhance model capacity.     Dense(512, activation='relu'),      ### Output layer with 5 units for the five weather classes using softmax probabilities.     Dense(5,activation='softmax') # values between 0 to 1  ])  ### Print the model summary to verify shapes and parameter counts. print(model.summary() )  ### Compile the model with Adam optimizer and categorical crossentropy for multi-class tasks. model.compile(optimizer='Adam', loss='categorical_crossentropy', metrics=['accuracy'])  ### Train the model for up to 100 epochs while validating on the held-out set and saving best checkpoints. history = model.fit(train_data_set , epochs=100 , verbose=1 , validation_data=val_data_set, callbacks=[best_model])  ### Extract training accuracy history for plotting. acc = history.history['accuracy'] ### Extract validation accuracy history for plotting. val_acc = history.history['val_accuracy'] ### Extract training loss history for plotting. loss = history.history['loss'] ### Extract validation loss history for plotting. val_loss = history.history['val_loss']  ### Create a range of epoch indices for the x-axis. epochs = range(len(acc))  #-> from 0 to the last epoch  ### Create a new figure to plot accuracy curves. fig = plt.figure(figsize = (14,7)) ### Plot training accuracy across epochs. plt.plot(epochs, acc, 'r' , label='Train accuracy') ### Plot validation accuracy across epochs. plt.plot(epochs, val_acc, 'b' , label='Validation accuracy') ### Label the x-axis with 'Epochs'. plt.xlabel('Epochs') ### Label the y-axis with 'Accuracy'. plt.ylabel('Accuracy') ### Set the title for the accuracy plot. plt.title('Train and validation accuracy') ### Add a legend to distinguish curves. plt.legend(loc='lower right') ### Render the accuracy plot to screen. plt.show()  ### Create a second figure to plot loss curves. fig = plt.figure(figsize = (14,7)) ### Plot training loss across epochs. plt.plot(epochs, loss, 'r' , label='Train loss') ### Plot validation loss across epochs. plt.plot(epochs, val_loss, 'b' , label='Validation loss') ### Label the x-axis with 'Epochs'. plt.xlabel('Epochs') ### Label the y-axis with 'Loss'. plt.ylabel('Loss') ### Set the title for the loss plot. plt.title('Train and validation loss') ### Add a legend to distinguish curves. plt.legend(loc='upper right') ### Render the loss plot to screen. plt.show()  ### Save the final trained model in case EarlyStopping did not trigger. model.save('C:/Python-cannot-upload-to-GitHub/Weather/bestWeatherModel2.h5')

You can find the full code here : https://ko-fi.com/s/efaafe52c5

Batch Prediction and Evaluation on Test Images

This section loads the saved model, preprocesses test images to a unified size, and runs batch predictions.
It prints true labels and predictions for quick inspection and saves the processed array for reuse.

The inference pipeline mirrors the training preprocessing for consistency.
Each image is loaded, resized to (256, 256), converted to a NumPy array, expanded to include a batch dimension, and normalized to [0, 1].
Keeping preprocessing identical to training is essential for faithful predict images keras behavior and reliable downstream metrics.

A list of test filenames is read from CSV to create a reproducible evaluation batch.
Vectorized concatenation produces a single tensor ImagesArray that flows efficiently through the model, leveraging GPU parallelism when available.
This is faster and less error-prone than per-image prediction loops when working with hundreds or thousands of samples.

Model outputs are class-probability vectors.
Using argmax converts these to discrete class indices that map cleanly to human-readable labels like “cloudy” or “rainy”.
Printing side-by-side true labels and predictions enables a quick qualitative sweep to catch obvious failure patterns before deeper analysis.

Saving the prepared batch as a .npy file accelerates iteration.
You can reuse the exact same tensor across different models or checkpoints without re-reading and re-processing images from disk.
This helps when comparing a baseline against a stronger augmentation regime or an alternative optimizer in TensorFlow image classification experiments.

For more rigorous evaluation, compute a confusion matrix, per-class precision/recall, and macro/micro F1 scores.
These metrics reveal whether the model systematically confuses visually similar conditions such as foggy vs. cloudy, guiding targeted data augmentation or class-specific sampling strategies.
Maintaining a clean separation between train, validation, and test ensures that reported improvements reflect genuine generalization rather than data leakage.

### Import Keras to load the trained model from disk. from keras.models import load_model ### Import NumPy for numerical operations and array manipulation. import numpy as np ### Import Keras image utilities for loading and converting images to arrays. from keras.preprocessing.image import load_img , img_to_array ### Import pandas for reading the CSV list of test image names. import pandas as pd  ### Define the model input width used during training. imgWidth= 256 ### Define the model input height used during training. imgHeight=256  ### Provide a human-readable list of class names for mapping predictions. classes = ["cloudy","foggy", "rainy", "shine", "sunrise"]  ### Load the best saved model from the training phase. model = load_model("C:/Python-cannot-upload-to-GitHub/Weather/bestWeatherModel.h5")  ### Print model summary to confirm the architecture is correctly loaded. print(model.summary() )  ### Define a preprocessing function that loads and normalizes a single image. def prepareImage(ImagePath):     ### Load the image from disk and resize to the model's expected dimensions.     image = load_img(ImagePath, target_size=(imgHeight,imgWidth))     ### Convert the PIL image into a NumPy array.     imgResult = img_to_array(image)     ### Add a batch dimension so the shape becomes (1, H, W, C).     imgResult = np.expand_dims(imgResult, axis = 0)     ### Normalize pixel values to [0,1] to match training rescale.     imgResult = imgResult / 255.     ### Return the prepared image tensor.     return imgResult  ### Define the folder that contains the test images. testImagesFolder = "C:/Python-cannot-upload-to-GitHub/Weather/Test" ### Read the CSV file listing the test image filenames. testImagesNamesDF = pd.read_csv("C:/Python-cannot-upload-to-GitHub/Weather/test.csv") ### Initialize a Python list to collect full image paths. testImagesList = []  ### Convert the 'Image_id' column to a Python list of filenames. testDFList = testImagesNamesDF['Image_id'].tolist()  ### Iterate over the filenames to build absolute test paths. for item in testDFList:     ### Concatenate folder path and filename.     tempName = testImagesFolder + "/" + str(item)     ### Append the full path into the list of test images.     testImagesList.append(tempName)  ### Print the list of test image paths for verification. print("The list of the images : ") print(testImagesList)  ### Prepare the first image and initialize the batch array. ImagesArray = prepareImage(testImagesList[0])  ### Iterate over the remaining images to append them into the batch. for imgName in testImagesList[1: ]:     ### Log which image is being processed for transparency.     print("preparing image : " + imgName)     ### Preprocess the current image.     processedImage = prepareImage(imgName)     ### Append along axis 0 to grow the batch.     ImagesArray = np.append(ImagesArray,processedImage,axis=0)  ### Print the final batch shape to confirm the dimensions are correct. print("Images shape: ") print(ImagesArray.shape)  ### Persist the prepared batch to disk for reproducibility and later reuse. np.save("C:/Python-cannot-upload-to-GitHub/Weather/ImagesArray.npy", ImagesArray)  ### Run batch predictions with a reasonable batch size for speed and memory balance. resultArray = model.predict(ImagesArray, batch_size=16, verbose=1) ### Convert softmax probabilities to class indices using argmax. answers = np.argmax(resultArray, axis = 1) ### Print the numeric class predictions for quick inspection. print("Answers : ") print(answers)  ### Extract the ground-truth labels from the CSV. yTrue = testImagesNamesDF['labels'] ### Assign the predicted indices for comparison. yPred = answers  ### Initialize a counter to track image indices. num = 0 ### Iterate over test images to print ground truth and predicted class names. for imgName in testImagesList:      ### Print a per-image line with the file path, true label, and predicted label.     print ("Image :" + imgName + "      True Value :" + classes[yTrue[num]] + "      Predictions: " + classes[yPred[num]] )      ### Increment the counter to move to the next image.     num = num + 1

You can find the full code here : https://ko-fi.com/s/efaafe52c5

Connect :

☕ Buy me a coffee — https://ko-fi.com/eranfeit

🖥️ Email : feitgemel@gmail.com

🌐 https://eranfeit.net

🤝 Fiverr : https://www.fiverr.com/s/mB3Pbb

Enjoy,

Eran

error: Content is protected !!
Eran Feit