Land Use and Land Cover Classification

Land use and land cover classification is a remote sensing task. It matters for agriculture, urban planning, environmental monitoring, and disaster response. In this project, I trained and evaluated CNNs that classify Sentinel-2 image patches from the RGB EuroSAT dataset into ten classes, from forests and rivers to highways and residential areas.

I started in a notebook on Google Colab, then turned the work into reusable scripts. The final models were trained on a rented NVIDIA RTX 6000-class GPU.

Dataset

EuroSAT RGB has 27,000 image patches of 64 × 64 pixels across ten classes. I split it into stratified train, validation, and test sets in a 70/15/15 ratio with a fixed seed, which leaves 4,050 images for the test set.

Random samples from each class:

A grid of random EuroSAT samples, one row per land-cover class

The classes are slightly imbalanced, with 2,000 to 3,000 images each, so training uses a class-weighted loss.

Number of images per class in EuroSAT

Per-class RGB pixel intensity distributions

Method

Both backbones start from ImageNet-pretrained weights in torchvision. Each one gets a new classifier head for the ten EuroSAT classes, and training runs in two phases:

Frozen backbone: train only the new head, for 10 epochs at a learning rate of 1e-3.
Fine-tuning: unfreeze the whole network and continue from the best phase-one weights, for 20 epochs at a learning rate of 1e-5.

Images are resized to 224 × 224, and training uses horizontal flips, random rotation, and color jitter. Training also uses the AdamW optimizer, class-weighted cross-entropy, ReduceLROnPlateau, early stopping on validation loss, mixed precision on CUDA, and checkpoint resume.

Results

The evaluation used the held-out test set of 4,050 images.

Model	Accuracy	Macro F1	Macro AUC	ms / image	Parameters
VGG16-BN	96.69%	0.9665	0.9993	139.74	134.3M
MobileNetV2	95.26%	0.9511	0.9986	27.97	2.9M

VGG16-BN is the most accurate model. MobileNetV2 is only 1.4 points behind, and it is about 46 times smaller and 5 times faster. That makes it the more practical choice when latency or memory matters.

Accuracy, F1, and latency comparison of VGG16-BN and MobileNetV2

VGG16-BN

VGG16-BN confusion matrix on the test set

VGG16-BN training and validation curves

The wrong predictions with the highest confidence are useful for error analysis. For VGG16-BN, all five are annual crops predicted as permanent crops, two classes that can look almost the same from above.

VGG16-BN’s highest-confidence misclassifications

MobileNetV2

MobileNetV2 confusion matrix on the test set

MobileNetV2 training and validation curves

MobileNetV2’s highest-confidence misclassifications

Stack