Random Forests and Living Trees

Conditional Cover Class
Bulrush (Reeds) – dense bulrush thickets	Bulrush thickets on Hasanska Street	The boundary between water and bulrush. A pedestrian bridge over the Saksahan River near Galileo Street	The bulrush thickets. A pedestrian bridge over the Saksahan River near Diderot Street	The forest with dense bushes. Near the car bridge on Sicheslavska Street
Burned Area – areas of vegetation burned by fires	No photos were available. I relied on oral reports from city residents.
Forest – dense tree plantations with bushes and grasses	The with dense bushes. Near the car bridge on Sicheslavska Street	The forest with dense bushes. Near for the Fadeev Street	The meadow in the dense forest. Near the Nikopol Highway Street	The forest with dense bushes. Near the Nikopol Highway Street
Lawn – areas of low grass	The green lawn near the City council building	The stadium of General School is close to 66 Vokzalna Street	The withered lawn on the Ukrainska Street	Lawns in the square near Tambovska Street
Park – sparse tree plantations without bushes	Park near the former cinema “Druzhba”. Magistralna Street	Yuri Gagarin Park	Park near the former cinema “Druzhba”. Magistralna Street	Park with dense trees near the Museum of Labor Glory PrJSC “Pivnichny GZK” Cherkasova Street
Solid Surface – any hard surface: roads, concrete, bare soils, rock outcrops, industrial sites (quarries, dumps, etc.)	Hardcover near the City council building	Wasteland and buildings near Nikopol Highway	The road along Prorizna Street	The industrial zone around the “Pivnichny” quarry
Village – a dense mixture of trees, grasses, hard surfaces, buildings, and structures. Such a mix of heterogeneous objects within one pixel of a satellite image leads to specific reflection of both optical and radar waves. The most understandable representative of this class is single-story residential buildings, although areas of the “Village” class are also present in the yards of high-rise buildings.	Village on Nesterova Street	View of the village on Khersonska Street	Village on Tambovska Street	Village on Zaporizka Street
Wasteland – areas of tall grass with scattered trees	Wasteland near the pumping station on Galileo Street	Meadow between oak groves. Kershenska Street	Wasteland near the industrial site of the “Oktiabrska” mine. Fadeev Street	Abandoned stadium (?) near Fadeev Street
Water – any open water surfaces (rivers, ponds, reservoirs, etc.)	Saksahan River near Fadeev Street	Boundary between water and bulrush. Pedestrian bridge over the Saksahan River near Galileo Street	Saksahan River. Near the car bridge on Sicheslavska Street	View of the flooded quarry near Rubizhna Street

Based on the reference sites, I formed training and test samples for territory classification. The total number of reference sites is 1070, covering 25258 pixels of the satellite image (one-pixel size is 10×10 m).

I formed training and test samples from individual pixels within the training areas of each conditional class. For the test sample, 1200 pixels of each class were selected using a random number generator. Given that the “Burned Area” class has only 871 pixels, “sampling with replacement” was applied when forming the training sample. Therefore, the test sample for this class did not contain unique pixels, which almost did not affect the final result. It is worth noting that sampling with replacement led to possible duplication of “training” and “test” pixels for other classes.

I used satellite images from the Sentinel-1A, Sentinel-2A, and Landsat-8 satellites for the study. The main image was from Sentinel-2A, dated August 8, 2017. Additionally, I used a Sentinel-1A image from August 7 and thermal infrared channels from a Landsat-8 image dated August 12, 2017. A near-infrared (vegetation) band image from Sentinel-2A, dated April 30, 2017, was added at the final stage. The spring image allowed for a more precise separation of bulrush thickets from parks and forests.

The city boundary of Kryvyi Rih was taken from the OpenStreetMap project, excluding the “Avangard” settlement. The classification of land cover types within the “Avangard” settlement was not performed, and its area was not included in the total area of the studied territory.

As the classification algorithm, I chose Random Forest from the corresponding library of the R programming language. The number of trees in the forest I grew is 350, with each tree based on 4 variables (individual satellite image bands). The classification accuracy on the training sample reached 96.22%, with an “Out of Bag Error” of only 3.78%. The test sample showed an overall classification accuracy of 93.95%, which is quite high for such studies, given the high detail of the classification map (10×10 m). However, the accuracy of detecting individual classes varies significantly. The most significant error is shown by the “Lawn” class (25.15%), followed by the “Park” (11.45%) and “Forest” (9.78%) classes. Due to classification errors there were some mistakes in the final classification map. For example, some sections of the eastern sides of Kryvyi Rih quarries were classified as “Water” due to being in shadow during optical and radar imaging. Single pixels within residential areas were also classified as “Water,” again due to shadows from buildings. The table below shows the results of the classification model test sample. The rows contain predicted class values, and the columns contain real class values. The error in detecting each class is defined as the ratio of “incorrect” to “correct” model responses.

I obtained a classification map based on the classification of the entire territory of Kryvyi Rih. With this map, we can count the number of pixels of each conditional class within Kryvyi Rih. Knowing the area of one pixel (100 sq.m), we can calculate the total area of each class. The table below shows the rounded areas of the classes.

Below is the overall view of the final classification map. The online version of the map can be viewed here.