// Earth Observation Network · Active

Land Type Classification with Sentinel-2 Satellite Images

20K Dataset Images
10Classes
91.80%Accuracy
13Spectral Bands

EuroSAT — Land Use
Classification

Our model is trained on EuroSAT, a benchmark dataset built entirely from Sentinel-2 satellite imagery captured over Europe. Each image is stored as a multi-band .tif file, preserving the full radiometric depth across all 13 spectral bands — from visible light through shortwave infrared.

The dataset contains 20,000 geo-referenced image patches, each 64×64 pixels, covering 10 distinct land-use and land-cover classes. It is perfectly balanced, with 2,000 samples per class, ensuring no category dominates training and that the model learns each terrain type with equal fidelity.

Working with raw .tif files allows us to retain all spectral information discarded by conventional RGB formats — enabling the classifier to detect subtle signatures invisible to the human eye, such as vegetation stress, soil moisture, and urban heat variance.

SOURCEESA SENTINEL-2
TOTAL IMAGES20,000 PATCHES
CLASS BALANCE2,000 PER CLASS
IMAGE FORMATGEOTIFF (.TIF)
PATCH SIZE64 × 64 PIXELS
SPECTRAL BANDS13 BANDS · 443–2190 NM

// 10 Land-Use Classes · 2,000 Images Each

🌾 Annual Crop
🌲 Forest
🌿 Herbaceous Vegetation
🛣️ Highway
🏭 Industrial
🌊 Pasture
🏘️ Permanent Crop
🏙️ Residential
🌊 River
🏞️ Sea & Lake
FILE FORMATGEOTIFF · MULTI-BAND · LOSSLESS
BAND RANGEB01–B13 · VISIBLE + NIR + SWIR
NORMALISATIONPER-BAND MIN-MAX · PRESERVED ACROSS SPLITS
TRAIN / VAL / TEST70% · 15% · 15%

// Methodology

Building a High-Accuracy
Classification Pipeline

01

// Pre-processing

Spectral Normalisation

Each .tif image carries 13 raw spectral bands with pixel values spanning different physical ranges depending on wavelength. Before any learning can occur, all bands are normalised per-channel using min-max scaling, squeezing every value into [0, 1] while preserving relative spectral contrast. Statistics are computed on the training split only and applied identically to validation and test sets, preventing any form of data leakage across partitions.

02

// Dimensionality Reduction

PCA — 8 Principal Components

With 13 correlated spectral bands, redundancy is inevitable — neighbouring infrared channels in particular share significant variance. Principal Component Analysis was applied across the flattened band space, retaining the 8 components that collectively explain the dominant variance in the data. This reduction compresses the input without discarding discriminative spectral signal, accelerates training, and helps regularise against overfitting on high-dimensional per-pixel features.

03

// Architecture Search

Model Selection — Four Candidates

Four pre-trained convolutional architectures were benchmarked under identical training conditions — same optimiser, same learning rate schedule, same augmentation policy — to isolate the effect of architecture alone. The candidates spanned a wide capacity range: AlexNet (lightweight, ~61M parameters, deep linear head), GoogLeNet (Inception modules, multi-scale receptive fields), ResNet-50 (residual skip connections, 50 layers), and EfficientNet (compound scaling, state-of-the-art efficiency). All models were fine-tuned from ImageNet weights with only the final classification head replaced for 10 EuroSAT classes.

04

// Result

AlexNet — Best Performing Model

Against expectation, AlexNet outperformed all three larger architectures on the EuroSAT benchmark. Its shallower depth and wider fully-connected layers proved well-suited to the relatively uniform 64×64 patch structure of the dataset, where the spatial patterns are coarser than natural ImageNet scenes. The winning configuration used RMSprop (lr = 0.0001, momentum = 0.9, batch size = 32) over 8 epochs, achieving a final validation accuracy of 91.80% — climbing steadily from 83.85% at epoch 1 to peak performance at epoch 8.

// AlexNet · Training Log · Best Trial

CONFIG LR 0.0001 · BATCH 32 · EPOCHS 8 · RMSprop · MOMENTUM 0.9
E1/8
83.85%
E2/8
89.78%
E3/8
89.90%
E4/8
85.33%
E5/8
89.97%
E6/8
91.35%
E7/8
91.60%
E8/8
★ BEST 91.80%

// Model Comparison · Same Hyperparameters

01
AlexNet

91.80% ★ Selected

Lightweight deep head, ideal for 64×64 coarse patch structure. Fast convergence, minimal overfitting.

02
🔀
GoogLeNet

Inception multi-scale modules. Strong on complex scenes, but over-parametrised for small uniform patches.

03
🔗
ResNet-50

Skip connections mitigate vanishing gradients across 50 layers. Competitive but slower to converge on this scale.

04
📐
EfficientNet

Compound scaling across depth, width, and resolution. Highly efficient on large inputs; overhead outweighed gains at 64px.

Mission Control.
Human Intelligence.

Behind every orbital data stream is a team of scientists, engineers, and visionaries who believe that seeing Earth more clearly makes protecting it possible.

CEO
Eng. Sohaila Mostafa
AI Engineer

Astrophysicist turned space entrepreneur. Former NASA Earth Science Division Director. 20 years in orbital mechanics and satellite design.

in
tw
CEO
Eng. Dina Mohsen
Data Analyst

Astrophysicist turned space entrepreneur. Former NASA Earth Science Division Director. 20 years in orbital mechanics and satellite design.

in
tw
SCIENCE
DR. LENA VASQUEZ
Head of Earth Science

Climate scientist and remote sensing expert. Her research shaped IPCC satellite data protocols. Leads our global environmental monitoring programs.

in
rs
DATA
JAMES HOLT
Head of AI & Data

Machine learning pioneer with a geospatial AI background. Built the algorithms turning raw satellite pixels into actionable intelligence within seconds of capture.

in
gh
OPS
NADIA AL-RASHID
Director of Operations

Former ESA mission controller. Oversees our 24/7 ground station network across 6 continents, ensuring zero-downtime data pipeline operations globally.

in
tw