Experimental statistics

Satellite images

tt

Illustration

Land cover mapping from satellite images

Methodology

Accurate and up-to-date land use and land cover maps are important in a wide range of areas such as urban and regional planning, disaster and hazard monitoring, natural resource and environmental management. With population, resource and environmental problems becoming increasingly prominent, an accurate and wide-ranging land use and land cover classification method based on remote sensing data can help to monitor many important large-scale issues.

For this task, we selected a deep learning (DL) model developed based on Corine Land Cover programme, which provides land cover/use status for 2018, and we refined the model using Sentinel-2 satellite images of the territory of Lithuania. The work was carried out using ArcGIS Pro software and Python programming language.

Data and data processing

The data used in this task are satellite images of the territory of Lithuania consisting of the full spectrum of Sentinel-2 bands – there are 13 Sentinel-2 bands in total, ranging from 10 to 60-meter pixel size. Blue (B2), green (B3), red (B4), and near-infrared (B8) channels have a 10-meter resolution; red edge (B5), the near-infrared NIR (B6, B7 and B8A), and short-wave infrared SWIR (B11 and B12) have a 20-meter resolution. Finally, coastal aerosol (B1) and cirrus band (B10) have a 60-meter pixel size. The satellite image sheets consisting of the above spectral bands were selected by reducing the cloud cover and extracting the spring–summer periods. Then these sheets were cut into 512x512 sub-sheets and the values of each band were normalised to the interval [0,1] to have a uniform view of signal dispersion. A total of 578 images were collected and labelled with over 40,000 unique features corresponding to the Corine classification.

Marking of features of territories

In remote sensing based on deep learning algorithms, the definition and acquisition of the training dataset are often critical issues. This is usually done manually: the territories corresponding to each class are labelled and submitted for training. In this work, the Georeference Base Cadastre of Lithuania (GRBC) – the national cadastre in which natural and anthropogenic objects of the land surface are registered – was used to address this issue. This Cadastre was equalised with the Corine land cover classification classes and a training dataset was created for the deep learning model. Moreover, since each class has a different number of samples, data augmentation was performed to match these classes, which is a widely known method to improve the model’s classification accuracy.


Figure 1. Conversion of satellite image to GRBC classes

Illustration

 

Refinement of the Corine classification model for Lithuania

The Convolutional Neural Network (CNN) is the most widely used DL neural network for image processing. CNN is a category of neural networks that has proven to be very effective in areas such as image recognition, image segmentation, classification, computer vision, and natural language processing. By using the generated training data, the basic Corine classification model, developed based on CNN, was improved for Lithuania’s land cover recognition. During the model training, an average accuracy of 88% was achieved in recognition of different classes.


Figure 2. Equivalents of GRBC and Corine classification

 

 

   Result

    This is an initial result and work on the model improvement is continued.