Integrated usage of historical geospatial data and modern satellite images reveal long-term land use/cover changes in Bursa/Turkey, 1858–2020

Data Used

We used cadastral maps from 1858 to reconstruct the LULC structure of Aksu and Kestel for the mid-nineteenth century. General Staff of the Ottoman Army produced these maps in 1:10,000 scale. These maps were one of the earliest attempts of creating cadastral maps in the Ottoman Empire. The images of historical maps scanned at 1270 dpi resolutions are provided by the Turkish Presidency State Archives of the Republic of Turkey – Department of Ottoman Archives (Map collection, HRT.h, 561–567). Individual tiles of cadastral maps are of a 1:2,000 scale. However, these maps are kept separated from their accompanying cadastral registers or documentation regarding their production process in the archives. There are no studies on the production of these cadastral maps, but few studies used them until now^35,36.

The LULC structures of Aksu and Kestel for the mid-twentieth century were generated using aerial photographs from June 23, 1955, with a scale of 1:30,000. These aerial photographs were captured by the US Navy Photographic Squadron VJ-62 (established on April 10, 1952, re-designated to VAP-62 on July 1956, and disestablished on October 15, 1969). The squadron conducted mapping and special photographic projects worldwide³⁷. Lastly, the VHR satellite images of WorldView-3 (WV-3) with 0.3 m of spatial resolution, collected on September 6, 2020, were used to show the up-to-date LULC patterns of Aksu and Kestel.

Methodology

Figure 2 shows the flowchart of steps followed in this study to detect the LULC changes. The workflow includes three phases: preprocessing, LULC mapping, and statistical analysis of LULC changes.

Data preprocessing

Orthorectification is the first important step in ensuring accurate spatial positioning among the multi-temporal and multi-source images, eliminating geometric distortions, and defining all data sets on a common projection system. To align the multi-modal geospatial datasets, we first performed the orthorectification of the satellite images and then we used the orthorectified satellite images as reference for the georeferencing of the cadastral maps and aerial photographs.

Satellite imagery orthorectification

We first pan-sharpened the WV-3 images by applying the PANSHARP2 algorithm³⁸ to fuse the panchromatic (PAN) image of 0.3 m spatial resolution with four multispectral bands (R, G, B, and near-infrared (NIR)) of 1.2 m. We then geometrically corrected the pan-sharpened (PSP) WV-3 imageries using an ALOS Global Digital Surface Model with a horizontal resolution of approximately 30 m (ALOS World 3D – 30 m), rational polynomial coefficients (RPC) file, and additional five ground control points (GCPs) for the refinement. As a geometric model, we used the RPC model with zero-order polynomial adjustment³⁹, and orthorectified images were registered to the Universal Traverse Mercator (UTM) Zone 35 N as the reference coordinate system.

Georeferencing of scanned cadastral maps and aerial photographs

We used orthorectified WV-3 imageries as a reference for the geometric correction of the historical cadastral maps and the aerial photographs. We selected the spline (triangulation) transformation, a rubber sheeting method, useful for local accuracy and requiring a minimum of 10 control points, as the transformation method to determine the correct map coordinate location for each cell in the historical maps and aerial photographs. The spline transformation provides superior accuracies for the geometric correction of the historical geospatial data^40,41.

The lack of topographic properties of planimetric features in the historical cadastral maps and the inherent distortions of the aerial photographs due to terrain and camera tilts causes difficulties in precise georeferencing of these data sets. It increases the number of required ground control points (GCPs) for optimal image rectification. Adequate and homogenously distributed GCPs, identified from cadastral maps and aerial photographs, can ensure precise spatial alignment among different geospatial data. The best locations for GCPs were border intersections of agricultural fields and roads. As for artificial objects, places of worship and schools, which are highly probable that have remained unchanged, are other optimal locations for GCPs to perform the accurate geometric correction. Figure 3 displays samples of GCPs selected from cadastral maps and aerial photographs. We obtained 2.11 m or better overall RMSE (Root Mean Square Error) values for the geometric correction of the historical maps and aerial photographs.

LULC classification scheme

We defined our classification scheme by analyzing the LULC classes distinguished in all three datasets (i.e., cadastral maps, aerial photographs, and WV-3 imageries). We used the classification scheme shown in Table 1. We also present codes and names of the land cover (LC) classes derived from Corine LC nomenclature⁴².

The legends provided on the historical cadastral maps of Aksu and Kestel delineate 15 LULC categories, which are: (1) buildings, (2) home gardens, (3) roads, (4) arable land, (5) gardens, (6) mulberry groves, (7) chestnut groves, (8) olive groves, (9) vegetable gardens, (10) forest, (11) courtyards, (12) vineyards, (13) arable fields, (14) cemeteries, (15) watercourses. Categorizing the land cover types of cadastral maps is limited with the classes available in the map legend. The legend of cadastral maps categorizes the forested area in one class named “forest”. Therefore, it was not possible to use third-level LC sub-categories in our classification schema for forest area which is separating forested areas into three subclasses (3.1.1, 3.1.2, and 3.1.3) according to the type of tree cover. Although some of the third-level LC sub-categories could be extracted from the cadastral map legend, it was not possible to extract all third level agricultural classes from single-band monochromatic aerial photographs. Although the spatial extent of fruit trees as a permanent crop could be determined from aerial photographs, it was not possible to classify these trees into different fruit types (e.g. 2.2.1 Vineyards, 2.2.2 Fruit trees and berry plantations, 2.2.3 Olive groves). Limitation on the number of forest classes is due to the historical cadastral map content; whereas limitation on the number of agricultural classes is mainly offset by the aerial photographs which have only one spectral band and we did not have any field survey or ancillary geographical data that could help the specific identification of fruit trees.

Our primary focus is to find out agricultural land abandonment, deforestation/afforestation, urbanization, and rural depopulation within the historical periods. Therefore, most of the second level LULC classes are sufficient for our purpose. LULC changes within the third class level such as the conversion of third level agriculture classes among each other were not aimed to analyze in this research. Our datasets allow us to use Level 3 CORINE classes for the artificial surfaces. These classes are useful to determine residential area implications of rural depopulation or urbanization, one of the focused transformation types for our analysis.

We re-organized and categorized the LULC types used in the cadastral maps, with minimum possible manipulation (only for 2.4 and 3.2 LC classes) according to the classification scheme, as shown in Table 2.

LULC mapping

After aligning all geospatial data, we used the georeferenced cadastral maps, aerial photographs, and satellite images for the LULC mapping. We set the spatial extent of the selected regions based on boundaries digitized from the cadastral maps of 1858. Then we detected historical LULC changes within these extents for all geospatial datasets covering 1900 ha and 830 ha of the Aksu and Kestel regions, respectively. Figures 4 and 5 show the selected extents from the historical maps, aerial photographs, and satellite images of the Kestel and Aksu sites, respectively.

Digitization of cadastral maps-1858 LULC maps

We visually interpreted and manually digitized the geographic features on the historical maps and created vector data for each class. The road features in cadastral maps lack topological properties. They also include spatial errors possibly generated in the process of surveying and map production. Therefore, we cross-checked digitized road segments by visual inspection of the road data of the aerial photographs from 1955. We then further modified road polygons to represent accurate road widths. Afterward, we categorized vectorized features of the cadastral maps into the LULC classes defined in Table 1. Finally, we created the vectorized 1858 LULC map. Figure 6 presents the vectorized 1858 cadastral maps of Aksu and Kestel.

Object-based image analysis of aerial photographs-1955 LULC maps

At the second stage of LULC mapping, we performed the segmentation and classification of the aerial photographs using an object-based approach for generating the 1955 LULC map. The object-based image analysis (OBIA) approach in LULC mapping provides advantages over the traditional per-pixel techniques such as higher classification accuracy, depicting more accurate LULC change, and differentiating extra LULC classes^33,43,44. We used the eCognition® software (Trimble Germany GmbH, Munich) to implement an object-based image analysis (OBIA). The OBIA approach contains two phases including the segmentation and classification phases that are performed to locate meaningful objects in an image and categorize the created objects, respectively.

Multiple ancillary datasets have been used to support different phases of OBIA. The Open Street Map (OSM) vector data, an open-source geospatial dataset (http://www.openstreetmap.org/), has been utilized as ancillary vector data in OBIA to improve the classification of the remotely sensed images. Sertel et al. (2018) used OSM as a thematic layer for road extraction⁷. Since there are several limitations in extracting the roads from aerial imagery, the OSM road network data could be useful. A majority of unpaved roads in single-band aerial photographs can easily be misclassified as homogeneous areas of arable lands. Precise detection of the roads from monoband aerial photographs without multi-spectral information is difficult. Therefore, we overlaid the OSM road network data with the aerial photographs to extract the revised aerial road vectors through visual interpretation and manual digitization.

We segmented the 1955 aerial photographs with the integration of 1858 LULC map produced from cadastral maps. We implemented the multi-resolution segmentation algorithm. In this segmentation method, a parameter called scale determines the size of resulting objects, and the shape and compactness parameters determine the boundaries of objects. The segmentation process of the aerial photographs was performed at multiple stages with various scale, shape, and compactness parameter values. At the initial stage, we segmented the regions according to the 1858 LULC map and we utilized large-scale parameters. The scale parameter was set to 100 and the shape parameter and the compactness were set as 0.7 and 0.3, respectively. At this stage, we focused on interpreting the objects that have not changed between 1858 and 1955. We classified the segments using the thematic layer attribute (LULC classes defined by the cadastral maps) with the highest coverage. Image objects in which the land surface has changed during 1858–1955 period were detected by visual interpretation and unclassified for further segmentation. We followed this approach to reduce the manual effort. We defined unchanged objects between 1858 and 1955 and assigned the same classes of 1858 LULC map to the objects in 1955 aerial photographs. We then segmented the remaining segments, the last time into smaller objects with the scale parameter set as 25, the shape parameter set as 0.2, and the compactness set as 0.8.

We classified the remaining unclassified objects through the development of rulesets. An object can be described by several possible features as explanatory variables which are provided by eCognition. In the classification ruleset, different features and parameters can be defined to describe and extract object classes of interest and thresholds for each feature can be defined by the trial-and-error method. We tested sets of variables for the classification of the monoband aerial photographs. Object features such as the mean value of the monoband, texture after Haralick, distance to neighbor objects, shape features (e.g., rectangular fit and asymmetry), and extent features (e.g., area and length/width) were the most useful alternatives. The classification process of the parcels of the aerial photographs with LULC change started with the classification of roads constructed between 1858 and 1955 by utilizing the aerial road map. The watercourse class was the most difficult to classify since shrubs or trees mostly covered the watercourses. These areas were misclassified as forest or agricultural land. Therefore, experts in historical map reading with local geographical information performed the detection and classification of the water course class and interpreted by the cadastral map (1858) and the google map (2020). After roads and watercourses, we classified forest and agricultural lands using the optimal thresholds for the brightness feature. We calculated the thresholds using the single band of the aerial photograph combined with the area and rectangular fit features. The heterogeneous agricultural areas class principally occupied by agriculture with significant areas of natural grass and trees within the same object are separated from the arable lands using the standard deviation of the digital number (DN) values of the aerial photographs. The texture feature helped classify the permanent crops. The brightness, shape, asymmetry, and distance to road class features were the best-performing ones for classifying the remaining artificial surfaces. The manual interpretation was performed for the classification of sub-classes of artificial surface class, including the continuous/discontinuous urban fabric, industrial, commercial, and transport units, mine, dump and construction sites, and artificial, non-agricultural vegetated areas. Since these land use classes contain one or more land cover and land use categories (e.g., artificial non-agriculture land or industrial or commercial units), finding the optimal threshold and exact feature for distinguishing the subclasses of artificial surfaces is difficult. Especially in the case of using the single-band aerial photographs, manual interpretation was required.

Object-based image analysis of satellite images-2020 LULC maps

We segmented WV-3 satellite images using multi-resolution segmentation algorithm and ancillary geographic data. Similar to the aerial road map, the road network of the study region in 2020, named, WV-3 road map, was extracted by overlaying the OSM road data with the WV-3 satellite image. In the segmentation process of the WV-3 image, we used the vector boundaries of the classified aerial photograph (the 1955 LULC map) and the WV-3 road map as ancillary thematic layers. We opted for the same segmentation and classification approach used for the aerial photographs for the WV-3 image.

Firstly, we segmented the satellite image into spectrally homogeneous objects using vector data of the 1955 LULC map by applying large-scale parameters. We implemented scale parameter values of 300, 200, 100, and 50 to find the optimal scale to classify objects that have not changed between 1955 and 2020. The best multi-resolution segmentation configuration was the scale of 100 and the shape and compactness parameters of 0.3 and 0.7, respectively. We classified the segments using the thematic layer attribute (LULC classes defined by the aerial maps) with the highest coverage. Segments with LULC change, e.g. the image objects in which the land surface has changed during 1955–2020 period were detected by visual interpretation and unclassified for further segmentation. As a result, we excluded the objects which were remained unchanged during 1955–2020 by assigning the prepared labels which were allocated in the previous step during the classification of 1955 aerial photographs. We then segmented the remaining objects into smaller objects to identify the changed areas in detail. At this step, the scale, shape, and compactness parameters were set as 25, 0.2, and 0.8, respectively.

Except for the additional sets of variables utilized to classify the WV-3 images, we applied the rule-set developed for the classification of the aerial photograph for the classification of the remaining objects of 2020 satellite images. The additional sets of variables include the mean of G, B, R, and NIR and two spectral indices, the Normalized Difference Water Index (NDWI), and the Normalized Difference Vegetation Index (NDVI). NDVI was calculated as the normalized difference of reflectance values in the red and NIR bands; whereas , NDWI was determined as the normalized difference of reflectance values of the green and NIR bands. Through the logical conditions, objects having specified values of NDVI and NDWI can be assigned to vegetation and water classes, respectively. The use of NDVI facilitated the delineation of terrains covered by vegetation and the NDWI improved the extraction of water bodies due to its ability to separate water and non-water objects. We separated different sub-classes of agricultural areas and forests by using optimal thresholds for NDVI which were defined by a trial and error method. Also we utilized assigning the optimal threshold to NDWI to separate water bodies from other land covers. In addition, the mean blue band layer was useful in classifying the artificial surfaces. We assessed the accuracy of each classification using error matrices (overall, user’s and producer’s accuracies, and Kappa statistics)^45,46.

Estimating LULC changes and LULC conversions

After the production of LULC maps of Aksu and Kestel for 1858, 1955, and 2020, the vector data of the LULC maps were used to quantify the LULC conversions for two different periods which are 1858–1955 and 1955–2020. To compare the LULC maps of study areas between two different dates of each study period, we provided detailed “from-to” LULC change information by calculating the LULC change transition matrix computed using overlay functions in ArcGIS.

We overlaid LULC maps of 1858 and 1955 and intersected the vector boundaries of the 1858 and 1955 LULC maps to determine the conversion types of LULC classes (from which class to which class). Similarly, to quantify the LULC changes between 1955 and 2020, we overlaid the 1955 and 2020 LULC maps. Then we created transition matrices and performed statistical analysis utilizing the matrices. Finally, we discussed the main LULC change types and the driving factors of the changes in the selected study areas.

Source: Ecology - nature.com