Satellite Image Land Use Classification using CNN (EuroSAT Dataset)

Land-use maps are normally drawn by analysts who inspect satellite imagery patch by patch, a process that is slow, subjective and hard to repeat over large areas or over time. Two neighbouring patches can look deceptively similar from orbit — pasture and herbaceous vegetation, or annual and permanent crops, often differ only in texture, field pattern and seasonal colour — so manual labelling does not scale and different labellers disagree. This project automates that first pass with a convolutional neural network trained on the EuroSAT benchmark: each Sentinel-2 patch is classified into one of 10 land-use classes, with per-class probabilities so uncertain patches can be flagged for human review instead of silently accepted. The full pipeline, from data loading to confusion-matrix evaluation, is documented so the student can explain every decision.

1. Dataset: The EuroSAT RGB dataset is loaded — Sentinel-2 patches of 64×64 pixels in 10 land-use classes — and split with stratification into training, validation and test subsets.
2. Preprocessing: Patches are normalised to the range [0, 1], and training patches receive light augmentation (flips and small rotations) because satellite views have no fixed orientation.
3. Training: The compact CNN — convolution blocks with ReLU and max-pooling, dropout, a dense layer and a softmax output over 10 classes — is trained with the Adam optimiser and categorical cross-entropy loss.
4. Evaluation: The held-out test subset is evaluated once, producing overall accuracy, per-class precision, recall and F1-score, and the full confusion matrix for the report.
5. Inference: A new patch is resized to 64×64 pixels, normalised exactly as in training, and passed through the saved model to obtain a probability for every class.
6. Output: The demo and report present the top predicted class, the complete probability distribution and the ground-truth label where available, so errors can be inspected patch by patch.

Tech stack:

• Python 3, TensorFlow/Keras
• EuroSAT dataset (Sentinel-2 imagery)
• NumPy, Matplotlib
• scikit-learn (metrics and splits)
• Jupyter Notebook (training and evaluation)
• HTML, CSS and JavaScript (interactive demo)

Dataset & model details

• Dataset: EuroSAT (Helber et al., 2019) — 27,000 labelled Sentinel-2 satellite image patches, each 64×64 pixels, in 10 land-use and land-cover classes with 2,000 to 3,000 images per class (design target split: stratified training, validation and test subsets). This project uses the RGB version of the dataset; the original release also provides 13 spectral bands. Classes: AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial, Pasture, PermanentCrop, Residential, River, SeaLake.
• Task: 10-class single-label image classification; input = one 64×64×3 RGB satellite patch, output = probability distribution over the 10 land-use classes, with the highest-probability class taken as the prediction.
• Model: Compact custom CNN: Conv2D (32 filters) → MaxPool → Conv2D (64 filters) → MaxPool → Conv2D (128 filters) → MaxPool → Dropout → Flatten → Dense → Dropout → Dense with softmax over 10 classes, trained with Adam and categorical cross-entropy. A transfer-learning baseline is discussed in the report as a comparison option, not as the shipped model.
• Metrics: Overall accuracy (design target only — see Specifications), per-class precision, recall and F1-score, and a 10×10 confusion matrix on the held-out test subset. No accuracy figure is claimed as measured until the training run for the order is executed and documented.

• [Compact custom CNN classifier] Three convolution–pooling blocks followed by dense layers and a softmax output, designed for small 64×64 patches so training stays feasible on a laptop or free GPU notebook.
• [EuroSAT dataset explorer] Browse example patches and counts for all 10 classes, so the class definitions and the visual differences between similar classes are understood before training begins.
• [Interactive classification demo] Select a sample patch and view the predicted land-use class with the full per-class probability distribution, honestly labelled as demo data in the web preview.
• [Confusion matrix and per-class report] Precision, recall and F1-score per class plus a 10×10 confusion matrix, so confusions such as Pasture versus HerbaceousVegetation can be discussed with evidence.
• [Reproducible training notebook] Data loading, stratified splitting, augmentation, model definition, training and evaluation in one notebook with every hyperparameter recorded and explained.
• [Training curves and checkpointing] Accuracy and loss plots for training versus validation, with the best weights checkpointed on validation accuracy rather than the final epoch.
• [Exported model with preprocessing code] Saved model weights bundled with the exact resizing and normalisation pipeline, so inference in the demo matches training-time preprocessing.

• Complete training and evaluation Jupyter notebook
• Trained CNN model file with preprocessing code
• Interactive classification and dataset-explorer web demo
• Confusion matrix, per-class metrics and training-curve plots
• Project report PDF (background, dataset, architecture, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (CNN design, EuroSAT classes, evaluation metrics)