Page 137 - Çevre Şehir ve İklim Dergisi İngilizce - Özel Sayı

Murat Arslan - Reyhan Çakir - İsra Akyazi
Nida Kumbasar - Ahmet Doğan - Emre Yavuz

2.3. Al-based Approaches and Land Cover Classification

When creating land cover classes, data sets are used in two ways: directly
and for use in models. Sentinel-2 satellite images are classified through
machine learning algorithms. The literature review revealed that DESA-based
solutions yield better results in image recognition and classificaiton studies
(Lecun, 1998, Salman, 2018, Sofu et al., 2020). Also, it is seen that U-Net
method yield better results in the production of land cover maps (Demir et
al.,2019). As a result, in the system created with a hierarchical structure, the
main classes in each tile are classified using the U-NET algorithm, while the
sub-classes are classified using the 3B-DESA algorithms.
9 satellite images from March and November each year are selected and
mosaicked based on Sentinel-2 tile and land cover classification was made on
grid-basis. Red, Green and Near-Infra-red bands are used in the classification.
Additionally, during land cover classification, a different algorithm is utilized
for each region.
The K-Fold cross-validation method (Aydemir et al., 2020) was used to
evaluate the classification models and measure their performance. In the
validation process, the vertical axis pixel count of the raster data produced
by the models was divided into five equal parts. Then, by random sampling
within each part, approximately two million samples from all main classes were
selected. Four out of these five parts were used in the training of 3B-DESA
classifier, and the remaining part was used as a validation data.
The kappa and overall accuracy results were calculated for the validated
region. This process was repeated to ensure that all parts underwent validation.
Later,a prediction map for the entire pilot region was created using the five
obtained classifier models The kappa and overall accuracy results for the entire
pilot region were calculated by comparing the prediction map with all the labels
from the data used in the training. This validation method is integrated into the
software, and the he K-fold (Chen et al., 2014) method is applied automatically
for each of the 128 tiles used to create the national map, generating error
matrices. The training data set is used for forming the classification model
while the validation data set is used for evaluating the classification model and
selecting the corresponding parameters. During analysis, the training data set
is viewed directly while the validation data set is viewed indirectly. The test data,
on the other hand, represents the validation set that has never been viewed
by classification model. While developing the classification model, training and
validation data sets are used together. The classification performance metrics
represent the prediction results achieved by the classifier model using the test
data. In this way, the performance of the classifier model is put to test with the
test data set that it has not previously viewed during training phase.

124 Journal of Environment, Urban and Climate