Deep Learning with Keras

Implememnation of Deep Learning models in keras.

Convolutional neural networks

SegNet

SegNet architecture (Badrinarayanan et al. 2015) for image segmentation.

DeepVel

DeepVel architecture (Asensio Ramos, Requerey & Vitas 2017) for regression.

Data

Segmentation

2D Semantic Labeling Contest

The semantic labeling contest of ISPRS provides two state-of-the-art airborne image datasets in Vaihingen and Potsdam, respectively for the detection of urban objects.

Six categories/classes have been defined:

1. Impervious surfaces (RGB: 255, 255, 255)
2. Building (RGB: 0, 0, 255)
3. Low vegetation (RGB: 0, 255, 255)
4. Tree (RGB: 0, 255, 0)
5. Car (RGB: 255, 255, 0)
6. Clutter/background (RGB: 255, 0, 0)

Sentinel-2 cropland mapping

Following the paper by [Belgiu & Csillik (2018)] (see also Hao et al. 2018)(https://www.sciencedirect.com/science/article/pii/S0034425717304686) we are going to train SegNet for the segmentation of the croplands. As an input we can use Sentinel-2 MSI multispectral data, and as an output crop types data classified by experts from the European Land Use and Coverage Area Frame Survey (LUCAS) and CropScape – Cropland Data Layer (CDL), respectively.

Datasets in Google Earth Engine:

• Sentinel-2 MSI: MultiSpectral Instrument, Level-1C
• USDA NASS Cropland Data Layers
• Canada AAFC Annual Crop Inventory

Plantations by type and species

• Plantations by type
• Plantations by species
• dataset

Global Annual Burned Area Mapping

Requirements

Prerequisites

• Keras 2.2.2
• TensorFlow 1.10.0

pip install --upgrade keras
pip install --upgrade tensorflow
pip install --upgrade tensorflow-gpu