deerta0103 / S1S2VHSR

Multi-sensor and Multi-scale data fusion through Convolutional Neural Networks

S1S2VHSR

Multi-sensor and Multi-scale data fusion for land cover mapping through Convolutional Neural Networks (CNN).

This repository supports a paper we have submitted to IEEE JSTARS. The study assesses the fusion of Sentinel-1 (S1) and Sentinel-2 (S2) satellite image time series in addition to a Very High Spatial Resolution (VHSR) SPOT image for land cover mapping via a 3 branch CNN architecture. The study was carried out on Reunion island.

Prerequisites

The code relies on Pyhton 3.7.6. The CNN models were implemented with Tensorflow 2.

Examples

Running the main architecture model

• fusion of S1, S2 and SPOT

python main.py s1_path s2_path ms_path pan_path gt_path

• See help for descriptions python main.py -h/--help

Running ablation models

• available cases: S1 and S2, S2 and SPOT, S1, S2 and SPOT

• example for S1 and S2 python main.py s1_path s2_path ms_path pan_path gt_path -s s1 s2

• example for S2 and SPOT python main.py s1_path s2_path ms_path pan_path gt_path -s s2 spot

Change some hyperparameters

• See help for all available options

• batch size: -bs default value 256

• learning rate: -lrdefault value 0.0001

• number of epochs: -ep default value 1000