Change detection for Finnish CORINE land cover classification

Transcription

1 Change detection for Finnish CORINE land cover classification Markus Törmä, Pekka Härmä, Suvi Hatunen, Riitta Teiniranta, Minna Kallio, Elise Järvenpää Finnish Environment Institute SYKE, Mechelininkatu 34, Helsinki, Finland ABSTRACT This paper describes the ideas, data and methods to produce Finnish Corine Land Cover 2006 (CLC2006) classification. This version is based on use of existing national GIS data and satellite images and their automated processing, instead of visual interpretation of satellite images. The main idea is that land use information is based on GIS datasets and land cover information interpretation of satellite images. Because Finland participated to CLC2000-project, also changes between years 2000 and 2006 are determined. Finnish approach is good example how national GIS data is used to produce data fulfilling European needs in bottom-up fashion. Keywords: Land cover, Land use, Change detection, Corine 1.Introduction The European Commission introduced the CORINE Programme in 1985 in order to gather information about the environment of the European Union. In order to determine and assess the effects of Community s environment policy, it is needed to have a proper understanding concerning the different features of the environment 1. CORINE land cover (CLC) classification is produced using satellite images. The mapping scale is 1: and mapping accuracy is at least 100 m. The minimum mapping unit is 25 hectares and minimum width of units is 100 m. Only area elements are classified. The classification nomenclature is hierarchical and contains five classes at the first level, 15 classes at the second level and 44 classes at the third level 1,2. There can also be national level 4 classes. Originally, the CLC classification was performed as visual interpretation of hardcopy printout of satellite images by overlaying a transparency on the printout, drawing polygons to transparency and digitizing drawn polygons 2. In order to update the CLC data European Environment Agency (EEA) and Joint Research Centre (JRC) launched the IMAGE2000 and CLC2000 project. The CLC2000 database is based on visual interpretation of Landsat ETM-images and ancillary data like existing maps. Interpretation is made using GISsoftware 2. As compared to the earlier version of CLC database, the updated version is more time-consistent, more accurate and costs are lower 3. The main outputs of the IMAGE2000 and CLC2000 projects at European level are 4 : 1. national and European wide satellite image mosaic for the year 2000 (IMAGE2000), 2. an updated national and European CORINE land cover classification for 2000 (CLC2000), and 3. database of land cover changes between 1990 and 2000 at national and European levels. Discussions during 2005 and 2006 between EU member countries, European parliament and the main EU institutions responsible for environmental policy, reporting and assessment have underlined an increasing need for factual and quantitative information on the state of the environment to be based on timely, quality assured data, in particular in land cover and use related issues. Therefore, EEA endorsed a proposal in summer 2005 to update CLC data together with high resolution land cover data as part of the implementation of the GMES fast track service on land monitoring. This service should provide on a regular basis core land cover and use change data that can be used by a wide range of downstream

2 services at European, national, regional and local level 5. This GMES land monitoring core service will deliver the following products: 1. Orthorectified satellite images for the reference year 2006 (+/1 year), 2. European mosaic based on satellite imagery called IMAGE2006, 3. Corine land cover changes , 4. Corine land cover classification 2006, 5. High resolution core land cover data for built-up areas including degree of soil sealing for year 2006, and 6. High resolution core land cover data for forest areas for year The work is distributed so that products 1 and 2 are made by ESA, data and service providers, products 3 and 4 by National Reference Centres of member states and products 5 and 6 by service providers with JRC. Up to now CORINE Land Cover mappings have been project based work organized by European Environment Institute. Each participating country has been responsible for data production in it own territory onwards CLC updates are part of European land monitoring which is organized within Global Monitoring of Environment and Security (GMES) programme by DG ENTR of the European Commission. GMES initial operation (GIO) has been accepted by European Parliament (regulation 911/2010) 6. Land monitoring service includes the continuity of Corine Land Cover with a new update for the reference year 2012, the production of 5 additional pan-european High Resolution Layers (HRL) and support to harmonization efforts of counties in order to improve synergies between pan-european and national land cover activities 6. High resolution layers include separate data layers describing forests, build-up areas, grasslands, wetlands and waters. GIO data is based on IMAGE2012 satellite image data, which is received using multiple satellite instruments. Data production includes both centralized and decentralized approach. Commercial service providers would make image post-processing and production of high resolution layers, individual countries CLC update and verification and enhancement of high resolution layers, and European Environment Agency EEA technical coordination and dissemination. All produced data sets should be available mid The aim of this paper is to document the semi-automatic change detection method used to produce changed areas between years 2000 and 2006, and present ideas to enhance the production of Finnish CLC2012. Applied semiautomated approach enables detailed, high resolution change monitoring also for national and regional applications. 2. Satellite images: IMAGE2000 and IMAGE2006 The production of Corine Land Cover classification is based on interpretation of satellite images. Satellite image coverage acquired for Corine 2000 is called IMAGE2000 and for Corine 2006 IMAGE2006. IMAGE2000 consists of Landsat ETM+ images, 36 images was needed to form almost cloud-free coverage of Finland 7. Most of the images were taken mid- or late-growing season, one third year 2000 and the rest during years 1999, 2001 and IMAGE2006 consists of IRS P6 LISS and Spot4 MS images 8. Because of smaller image size, number of images increased lot (80 IRS and 51 Spot) even the coverage had more gaps due to cloud cover. There were two coverage, one for mid-summer and another for spring. Only mid-summer coverage of IMAGE2006 was used to produce Finnish Corine Most of the images were taken year 2006, but 23 images were taken during 2005 and 10 during This illustrates the difficulty to obtain cloud free satellite image coverage of Finland during one summer using one or two instruments. The orthocorrection was made for images. Cloud and shadow masking was performed using visual interpretation. Atmospheric correction was made using method developed by VTT Technical Research Center of Finland in case of IMAGE2000 and using ATCOR2 of Erdas Imagine in case of IMAGE2006. Topographic correction was made at Northern Finland using Ekstrand method in case of IMAGE2000 and semi-empirical correction in case of IMAGE2006. Finally, mosaics for different vegetation zones and whole Finland were made 7,8.

3 Because there were quite a lot gaps in IMAGE2006 mosaic, SYKE received set of DMC-satellite images 9, acquired by SLIM-6 instrument of BEIJING-1 mission. 13 images were taken during summer Images have three channels: Green, Red and NIR corresponding to ETM+ channels 2, 3 and 4. Good property of these images is large image size (swath width about 320 km), drawbacks are small number of channels and worse spatial resolution that in case of ETM+, IRS or Spot. Clouds and their shadows were removed using visual interpretation and images were resampled to 25 m pixel size. These images were used for change detection in gaps of IMAGE Corine Land Cover classification This section describes how CLC2006 should be made according to the guidelines of European Environment Agency, and how Finnish approach is different. 3.1 EEA guidelines for CLC2006 According to EEA guidelines, the Corine Land Cover 2006 database is updated using "change mapping first" approach. This means that in order to produce CLC2006 database, land cover changes are interpreted first and then combined with CLC2000 database from which found errors have been corrected. Then database is generalized so that the minimum mapping unit would be 25 ha 10. The aim of producing CLC-changes is to have European coverage of real land cover changes that are larger than 5 ha, wider than 100 m, occurred between years 200 and 2006, and are detectable on satellite images. Change mapping is carried out by visual comparison of CLC2000 vector data, and IMAGE2000 and IMAGE2006 satellite imagery and subsequent direct delineation of change polygons. Delineation of changes must be based on CLC2000 polygons, so that the outline of the change polygon exactly fits CLC2000 boundaries. The use of ancillary data like topographic maps is recommended Finnish version of CLC The production of Finnish CLC database is based on automated interpretation of satellite images and data integration with existing digital map data, as illustrated in figures 1 and 2 7. Map data provides information describing land use and soils, and satellite images provide information about land cover and are used to update map data. Continuous land cover variables are transformed into discrete CLC classes using thresholds of these variables according to class descriptions in CLC nomenclature. Figure 1. The main data sources for main CLC-classes in case of Finnish CLC2000.

4 Figure 2. Production flow of CLC2000 in Finland. The same approach which was applied in CLC2000 project was repeated in CLC2006 project in order to produce land use and cover of Finland year Automated interpretation of satellite images and data integration with existing digital map data was applied. Additionally, specific classes were interpreted manually with the aid of IMAGE2006 and ancillary data. In order to detect changes between 2000 and 2006 two approaches were combined (figure 3): 1. the differences between high resolution land cover data sets 2000 and 2006 were evaluated together with the 2. changes detected using satellite data only i.e. IMAGE2000 and IMAGE2006. Land cover data and changes were produced in the original resolution of satellite data. This enables production also more detailed information for national use. This was necessary in order to get national funding for the work and access to the national LC databases. An automated generalization procedure from pixel based changes into polygons with MMU of 5 hectares was developed and applied GIS data The existing digital map data forms a base for land use data 7,8. The databases created by other data providers have been purchased and used in the different phases of the production. Input data required some preprocessing, in order that it was possible to use them in production of CLC2006: Figure 3. Production of Finnish CORINE land cover 2006 and changes

5 Housing and dwelling register: Point data corresponding to year 2006 was rasterized to 25 m grid, bufferized 25, 50 or 75 m according to the use of building and classified to CLC-classes. Bogs and open rocks of Topographic database: These were rasterized to 25 m grid. Database covered whole Finland 2006, Northern Lapland was missing from year 2000 data. Soil extraction sites of MOTTO-database: Sites with missing information were deleted. Some new sites were acquired from regional Environment Centres. Data was rasterized. Digiroad: Large roads and streets were picked from year 2005 database and vector data was rasterized. Agricultural lands were defined using Finnish Land Parcel Identification System (FLPIS) and rasterized to 25 m grid. Water areas: Lakes and sea were rasterized to 25 m grid. Rivers were rasterized to 10 m grid, generalized and transformed to 25 m grid. Very small islands (<2 pixels) were removed. The coverage and modernity of input data was increased by visual and semi-automatic interpretation of satellite images. Target areas were race and golf courses, airports, ports, mining areas, dump and construction sites, and peat production areas. These areas were rasterized to 25 m grid. Post-processing steps contained enhancements to original data: Bogs of Topographic database: if tree crown coverage of bog was less than 10%, then it was classsified as open bog. Soil extraction and dump sites, peat production areas and urban areas were updated using satellite images. Bare soil and built areas were interpreted from satellite images using thresholded NDVI vegetation index-images. Interpreted areas were classified to Soil extraction and dump sites, peat production areas or urban areas, if some neighboring polygon belonged to these classes. Agricultural areas with missing information were classified according to SLICES 2005 classification and forest interpretation. Agricultural fields and pastures with high tree cover were classified as forest Interpretation of forest variables Interpretation of land cover in forests was made at Finnish Forest Research Institute Metla using field plot data of the Finnish National Forest Inventory and IMAGE2006. It included prediction of three variables at pixel level: the actual canopy cover of the trees (%), the canopy cover of the broad-leaved trees (%), and mean height of the trees (dm). The training data for the pixel level prediction of the variables consist of the field plots of the 10th National Forest Inventory of Finland (NFI10) measured during , and in the very northernmost Finland NFI9 field plot data measured in The major changes, like regeneration cutting and significant damages (e.g., storm damages) between the field measurement date and image acquisition date were identified as thoroughly as possible and identified filed plots were removed from the training data. Also, similar major changes in the field data after the image acquisition until the field measurements were made were identified and these plots were removed from the training data. The total number of the plots used in estimation was about The predictions were calculated for the entire area of the country, independently of the land use and land cover class. The purpose was to guarantee that the predictions were made for all pixels outside of forestry land minus forest roads. The employed method is similar to that in use in the operative multi-source forest inventory of Finland and is called improved k-nn prediction. It is a version of non-parametric k-nn method 11,12,13. The validation was conducted during the prediction phase using one-leave-out procedure and comparing the pixel level predictions, as well as comparing area estimates to those based on field plot data only. Image data was Finnish IMAGE2006. Interpretation areas, typically individual satellite images, were merged to one nationwide mosaic. Continuous land cover variables were classified into discrete CLC classes by thresholding and combining these variables according to class descriptions in CLC nomenclature 1. However in Northern boreal zone in Finland modified definitions for forest classes were applied. Areas with crown coverage of 15-30% (tree height > 5 meters) were classified as forest. If standard CLC rules (CC > 30%) would have been used in northern Finland, a large number of sparse forests due to poor natural circumstances would have been omitted. Modification of the threshold value for crown cover was possible since crown cover was interpreted as a continuous variable.

6 3.2.3 Classification of Northern Finland In the northernmost Finland and especially in the mountainous areas the density and characteristics of vegetation is regulated by natural factors. In these areas, especially above tree line, forest variables only do not describe the land cover as defined in CLC nomenclature 1. Thus information also on ground vegetation and soils are needed in order to define CLC classes. These data were interpreted using decision tree approach (See5 by RuleQuest24) with the aid of IMAGE2006, biotope maps and national GIS data 14. One of the benefits of this kind of classifier is that variables can be continuous like satellite images or estimation results, or categorical like map layers or previous classification results. Following features were used in classification: Image channels: Atmospherically and topographically corrected IMAGE2006 image channels. Normalized Difference Indices of image channels: Six NDIs computed as ( ChA ChB ) / ( ChA + ChB ). MODIS NDVI: number of weeks the long-term MODIS NDVI is greater than 0.5 per year. DEM: Height from sea level, surface slope and aspect class. Forest estimates: tree height, crown cover and crown cover of deciduous trees. Soil: proportion of bogs, open rock, boulders and sand per hectare based on Topographic Database. Forest boundary mask. Classifications were made using individual images and the classification results were mosaicked according to highest classification confidence value provided by decision tree classifier. This interpretation was used in those areas of Northern Finland where the likelihood of pine forest is less than 70%. This likelihood is modeled based on average temperature sum and Biotope maps at Finnish Forest Research Institute Visual and semi-automatic interpretation of specific classes Some classes could not be produced from existing digital map data or interpreted using automatic methods. These classes were produced manually using visual interpretation with the aid of IMAGE2006, orthophotos, topographic maps and registers. These areas included large construction sites, golf courses, trotting-tracks, motor vehicle tracks, airports, harbors and dump sites. The digital map data does not always correspond with the situation visible in IMAGE2006. This is partly due to the mismatch of dates in IMAGE2006, IMAGE2000 and national data sets. Image data are acquired during 3-4 years, but the national data sets are extracted from data providers in the nominal year of CLC inventories i.e and Temporal mismatch of even 2 years exists. This led to the need to modify specific elements of national data sets using semiautomatic methods. The areal extent of certain build-up areas like industrial and commercial units and mineral extraction sites were delineated by combining automatically derived bare land mask with registers. Bare ground was estimated by density slicing of IMAGE2006 NDVI mosaic. Significant extensions of mineral extraction sites were also verified visually. Peat production areas were updated by combining visual interpretation of IMAGE2006 with SLICES classification and Topographic databases Construction of Finnish Corine Land Cover database EU version EU version of Finnish CLC was made so that correction mask of CLC2000 (raster with 25 m pixel size) and EU version of changes (5 ha MMU) were merged with national CLC2000 (raster with 25 m pixel size). The result, raster data with 25 m pixel size was generalized to European version which is vector data with 25 ha minimum mapping unit using generalization macros of CLC2000. This automatic generalization procedure is based on Arc/Info raster and vector operations and has following phases: Simplification of the input data, elimination of narrow linear features and single pixels Aggregation of the parcels to reach 25 ha inside each level 1 CORINE class Dissolving remaining small parcels to the most appropriate neighboring area according to the priority list Smoothing the boundaries of the parcels Merging working units to one single nationwide database

7 Vectorizing the raster database and final checks Heterogeneous classes like complex cultivation are produced during the generalization process Construction of Finnish Corine Land Cover database national version Finnish national CLC2006 was made completely differently than EU CLC2006. The aim was to make the best possible national CLC2006 by using the most correct and up-to-date information. This means that national CLC2000 and CLC2006 are not directly comparable because of differences in production methods and input data, and therefore if Finnish national CLC2000 and CLC2006 are directly compared the resulting changes are not real changes. Separate national CLC change-product contains detected real changes. Input data were preprocessed and post-processed as previously discussed, and merged to form data layers approximately corresponding to CLC level-1 classes. The order of merging these data layers was: visual interpretations of urban areas, wetlands, water, other built-up areas, agricultural areas, open rocks and sand, and forests and seminatural areas on mineral soil. This priority order was due to the quality, accuracy and age of input data, and experiences from CLC2000- project. Technical differences between Finnish national CLC2000 and CLC2006 are due to e.g. differences in input data like Topographic database covers now whole Finland. Forests estimates are now based on in-situ and interpretation methods of National Forest Inventory of Finnish Forest Research Institute. Use of new databases like Digiroad and Land Parcel Identification System. The gaps of CLC2006 due to clouds and their shadows of IMAGE2006 were filled using CLC2000, unless there was other data to fill them. Metadata was created during data merging phase. Source-metadata describes which input datasets have used to classify that particular pixel. Age-metadata describes that age of input datasets for each pixel. 4. CLC change detection in Finland Finnish approach to CLC change detection is different than the guidelines of EEA 10. The changes are defined by combining the areal extent of detected changes in image-to-image comparison and the thematic content of CLC2000 and CLC2006 classifications. Some specific changes, in artificial surfaces, arable lands and regrowth in forests, are based directly to these classifications, since they are difficult to detect using image-to-image comparison. 4.1 Image-to-image changes The applied change detection method needs to be fast and as automatic as possible due to large number of images. Simple image differencing using Red and NIR-channels was used. Pixel was defined as changed if the difference between images was larger that two times the standard deviation of differences. Red channel was used because it is sensitive to changes in the biomass of coniferous vegetation and NIR is sensitive to chlorophyll content. The difference images provide also information about the type of change i.e. loss or increase of vegetation as illustrated in figure 4. Since IMAGE2006 were resampled to 20 m pixel size and IMAGE m, the IMAGE2006 mosaics were resampled to 25 m pixel using cubic convolution interpolation method. In case of mismatch between images, IMAGE2006 image was georeferenced using Erdas Imagine Autosync and IMAGE2000 mosaic as base image. This was necessary since geometrical mismatch of images causes false changes in automatic procedures.

8 Figure 4. An example of image-to-image change detection, IMAGE2000 on the left and IMAGE2006 on the right. Colors of change polygons represent the type of change: red means decrease in red channel, blue increase in red channel, green decrease in NIR channel and cyan increase in NIR channel. 4.2 Changes in forest Changes within forest i.e. forest cuttings and re-growth cover most of the land cover changes in Finland. This is due high proportion of forests (over 70 % of the territory) and intensive utilization of forest resources. Forest cuttings can be interpreted with high accuracy by comparing multitemporal images, since the clearance of forest causes significant increase of reflectance values e.g. in red band of satellite data. Difference images of red bands were merely used in delineation of forest cuttings. The results were verified by comparing the total area of interpreted forest cuttings with official forest statistics regionally, which are produced annually by Finnish Forest Research Institute. The verification was made using the national version of CLC forest data, which is in the original resolution of satellite images. This was necessary since the European version of CLC change data is highly generalized with MMU of 5 ha. The area of forest cuttings is usually less than 2 ha in Finland. Re-growth in forests is problematic to delineate, since re-growth is a slow process and difficult to detect using satellite data only. Silvicultural measures make the situation even more complex, since especially young forests are managed frequently; cleaning and thinning of sapling stands, where especially deciduous trees and scrubs are removed from young coniferous forests. Re-growth areas were detected by comparing estimated forest parameters, which describe tree height and density. Also IMAGE2000 and IMAGE2006 data sets were used together with soil type, vegetation zone, historical land cover and regional forest statistics in change detection process (see figure 5). Improbable or impossible changes were further checked and a correction layer for the old CLC2000 data was produced. 4.3 Allowed and not-allowed changes Change information based on direct comparison of CLC2000 and CLC2006 classifications includes also false changes. This is due to the fact that the data collection methods of national data sets are under constant development. This leads to false changes, which are merely corrections to the previous data sets. Additionally, the data source and interpretation method of some land cover themes, like forest information have been changed since year In order to find real changes only, a look-up table is defined where impossible or highly improbable changes are either directly rejected or verified. Information based on image-to-image comparison was utilized as much as possible in cases where the change should be visible in the satellite data. Changes at bog areas were not classified real changes because there can be quite a

9 lot seasonal changes in water and moisture level, unless change was at peat production area or change from bog to agricultural area. Table 1 represents the allowed (x) and not-allowed (-) changes for EU version of Finnish CLC changes. Blue x means that change is allowed if it is supported by estimated forest growth, and red x means that change is allowed if it is supported by image-to-image changes. Boxed x means that there is that particular change in change data and pink box means that particular change exist in final, generalized EU CLC-change database (5 ha MMU). In a case that there is - with pink box, means that the change is due to generalization process. The amount of these kinds of changes is small. 4.4 Correction of CLC2000 While evaluating detected land cover changes between years 2000 and 2006, some changes were regarded as errors in old land cover data i.e. in CLC2000. These were due to the improvements in input land cover data and misinterpretations during generation of CLC2000. Additionally, during manual delineation of specific land cover categories some corrections were done into CLC2000 data. All corrections were combined into one layer, which was merged with old CLC2000 data into revised CLC2000 data set. Improvements in input data included: Topographic database, which is maintained by National Land Survey, covers now the whole of Finland, which was not the case during production of CLC2000. This improved the delineation of peatbogs and bare rocks in the northernmost Finland. The position of buildings has been checked in the building and dwelling database. The interpretation of forest variables was made by the Finnish Forest Research Institute using their expertise and in-situ measurements. Agricultural land parcel database produced by Ministry of agriculture and forestry was available including information describing geometry and species in agricultural areas. The active fields in the year 2000 where checked with the data of Finnish Land Parcel Identification System and Farm Register. Some associated dump areas of mines were reclassified from mines to dump sites. Figure 5. Detection of forestry changes.

10 Table 1. The allowed (x) and not-allowed (-) changes, EU version of Finnish CLC changes. 4.5 Production of CLC changes After combination of image-based changes together with CLC2000 and CLC2006 classifications a working version of land cover changes was available in raster format with 25 meter resolution, which was generalized in two ways: EU version: Changes were determined using CLC level-3 classes using MMU 5 ha and minimum width 100 m. National version: Changes were determined using CLC level-4 (level-2 for forest classes, i.e. tree species changes were omitted) classes using MMU of forest changes 1 ha and other changes 0.5 ha. 5. Results Tables 2 and 3 represent the change matrices of Finnish national change product (1 ha MMU) and EU change product (5 ha MMU), respectively. The total amount of changes is km 2 in national product and 7036 km 2 (about 2.1% of Finnish territory) in EU product. The majority of Finnish landscape is covered by forests. When looking EU changes, the most of the changes in Finland are due to forest management: forest cuttings and re-growth make 91% of all area of changes, cuttings 50% and re-growth 41%. The clearing of new agricultural land makes almost 7% of changes, and the enlargement of build-up areas covers about 1.5% of changes.

11 Table 2. Change matrix of Finnish CLC level-1 changes with 1 ha MMU, areas in km 2 on the left and percentages on the right. L1 L2 L3 L4 L5 L1% L2% L3% L4% L5% CLC CLC L1 L1% CLC CLC L2 L2% CLC CLC L3 L3% CLC CLC L4 L4% CLC00 L CLC00 L5% Table 3. Change matrix of Finnish CLC level-1 changes with 5 ha MMU, areas in km 2 on the left and percentages on the right. L1 L2 L3 L4 L5 L1% L2% L3% L4% L5% CLC CLC L1 L1% CLC CLC L2 L2% CLC CLC L3 L3% CLC CLC L4 L4% CLC00 L CLC00 L5% Table 4. Land cover flows describing the major land use processes in Finland, computed using national (CLC 1 ha) and EU changes (CLC 5 ha). LCF Land Cover Flows - level 1 CLC 5 ha [km²] CLC 1 ha [km²] CLC 5 ha [%] CLC 1 ha [%] Lcf1 Urban land management 13,2 10,8 0,2 0,1 Lcf2 Urban residential sprawl 34,4 148,5 0,5 1,0 Lcf3 Sprawl of economic sites and infrastructures 79,4 148,9 1,1 1,1 Lcf5 Conversion from forested & natural land to agriculture 471,6 725,5 6,7 5,1 Lcf6 Withdrawal of farming 9,8 257,3 0,1 1,8 Lcf7 Forests creation and management 6381, ,4 90,7 90,0 Lcf9 Changes of Land Cover due to natural and multiple causes 46,0 122,6 0,7 0,9 Table 4 represents the land cover flows describing the major land use processes in Finland. It can be noticed, that the differences between national and EU version of changes are quite small. The largest differences are with urban residential sprawl and withdrawal from farming, in which cases the proportion of changes is larger in national product meaning that typically these change areas are quite small, less that 5 ha MMU of EU product. The proportion of conversion from natural land to agriculture increases in EU changes, meaning that usually these areas are quite large, over 5 ha MMU. 6. Validation The technical team (TT) of European Environment Agency EEA visited Finland twice 15. The 3rd verification was completed by s. The first visit (spring 2008) was a combined training and validation of preliminary results in two pilot areas. TT visited the second time early June 2009 and a sample of Finnish territory was selected and validated from change data covering whole Finland. After taking into account as much as possible the suggestions made by TT, a new subset of data was validated remotely September These data were accepted by TT and the final version of CLC data was delivered mid-october 2009.

12 Figure 6. Forest changes in Finland according to national statistics (cuttings) and CLC2006 project (cuttings and re-growth) at different levels of generalization. CLC changes were compared with national statistics. Finnish Forest Research Institute Metla collects and publishes statistics describing completed forest management practices on annual basis for different parts of Finland. Collected data includes forest cuttings of different type (clear cuttings, thinnings etc). The total area and regional distribution of forest cuttings according to statistics were compared with detected forest cuttings of CLC2006. The large minimum mapping unit (5 and 25 ha) defined in CLC data specifications makes it difficult to compare national statistics with CLC results. The correspondence of detected forest cuttings in CLC2006 with statistics depends on the level of generalization. Since the average size of forest management unit varies between 1-3 hectares in different parts of Finland, change data with MMU of 5 ha underestimates the amount of changes. Since forest changes were first detected in the resolution of satellite data i.e. 25 meter raster, it was possible to generalize forest changes during the production process with different MMUs (1ha, 3.7ha and 5 ha were tested). Figure 6 illustrates that forest change data with MMU of 1 ha seems corresponds well with national statistics (underestimation by 3.1%). 7. Enhancements for CLC2012 Detection of re-growth of forests was problematic in CLC2006 project using automated methods with multitemporal satellite data. Interpretation process needs to be developed in order to be more objective and accurate. The idea would be to model forest growth using forest models, which use parameters like age of trees (or date of forest clearance), climate conditions (heat summation), site type and quality and elevation. Many of these parameters are available in grid format or can be approximated using existing GIS data sets. Availability of multitemporal, high resolution land cover mappings enables estimation of the date of forest clearance in Finland. The first land cover mapping was completed late 1980's, thereafter four mappings have been produced including two CLC mappings. This approach will be tested in the future CLC updates in order to enhance and focus change detection in forest using multitemporal satellite data. One of the difficulties has been the enlargement of point data. In other words, the knowledge about land cover or use is stored as point in database and its areal extent is unknown. So far, these sites have been expanded using NDVI-value but the results have not been that good. For example, MOTTO-database has coordinates for soil extraction sites, like right dot in figure 7 is sand pit and left dot bedrock crushing site. The example in figure 7 is made by segmenting IRS P6 LISS-image using ecognition-segmentation software, intersecting segments with MOTTO-points, and deciding which segments belong to soil extraction sites using NDVI-value. Black indicates segments with -0.1 < NDVI < 0.0, blue segments 0.0 < NDVI 0.2 and red segments 0.2 < NDVI < 0.3. In seems that in this case merging black and blue

13 segments gives quite good borders for sand pit. Also, it seems that bedrock crushing site (dot on left) has not been opened because the NDVI of that area is quite high. One of the new features of CLC2012 is high resolution layers. So far, the most promising high resolution layer is soil sealing which could be used to enhance the interpretation of urban areas. The coordinates of building are acquired from Building and Dwelling Register, but Soil sealing-layer could be used to delineate large buildings and parking lots. The importance of other high resolution layers is smaller but national CLC2012 could be used to validate them. Dividing wetlands in refined national sub-classes according to their characteristics has been demanded by users. This will be tested by using multi-temporal image coverage for mid-summer and spring and up-to-date HR or aerial imagery. 8. Conclusions Acquisition of a cloudfree multitemporal (spring and summer) HR satellite data coverage over Finland every 3 years is challenging with present instruments. Acquisition of IMAGE2000 data (mid-summer coverage only) took 4 years ( ) and acquisition of IMAGE2006 took 3 years. In IMAGE2006 mid-summer coverage there are missing data or cloudy areas covering 4.5 % of the total land area of Finland. In the spring coverage the corresponding figure is even 13.5 %. Additionally, some spring images are received in the snow-covered period. The Finnish method in CLC data production is based on combined usage of satellite data and existing national land cover data sets. The dates of national data and satellite images are different since national data sets are extracted from databases in the nominal year of CLC mappings and images are received within 3 years around these nominal years. This leads to the need to extra modification of specific elements of national data sets using semi-automatic methods together with image data, which working phase produces significant costs. Figure 7. Detection of borders of sand pit using segmented satellite image.

14 The validation CLC-changes by EEA Technical Team is based on visual interpretation of satellite data together with ancillary map data. Acceptance of land cover changes which cause only a minor visible change into multitemporal satellite data is problematic, which may lead to bias in change statistics. Well-visible changes like forest cuttings and clearance of new agricultural areas will be well mapped and accepted, while gradual processes like forest re-growth and forestation of abandoned agricultural areas can be underestimated. Since these gradual changes are often spatially unhomogenous and scattered, they are more often generalized away from the database compared to spatially homogenous changes. Thus, also a quantitative validation of CLC changes is needed. The production and updates of European land cover data is now based on projects launched and partially funded by EEA. Since also national funding and resources are requested, project-based organization of operational land monitoring is problematic. National commitment and organization of work takes time and causes delays in the data production. In the future agreements between EU and member states should be discussed and agreed on country level is order guarantee both national and EU resources and national organization of work (cooperation) on long term basis. If data production of European land cover is based on coordinated and long term programme, data needs of EU can be also taken into account in national monitoring programmes. All this would enable faster production of data and availability of national expertise and data sources, which guarantees also high data quality and national usage of produced data. References [1] Heymann, Y. Steenmans, Ch., Croissille, G., Bossard, M., [Corine Land Cover Technical Guide], Office for Official Publications of the European Communities, Luxembourg, 163 (1994). [2] Bossard, M., Feranec, J., Otahel, J., [CORINE Land Cover Technical Guide Addendum 2000], European Environment Agency Technical Report 40, 105 (2000). [3] Büttner, G., Feranec, J., Jaffrain, G., [Corine Land Cover Update Technical guidelines], European Environment Agency Technical Report 89, 56 (2002). [4] Krynitz, M., [Land Cover Annual Topic Update 2000], European Environment Agency Topic report 4, 28 (2001). [5] EEA, [EEA Project Implementation Plan GMES Land FTS ], European Environment Agency AMP AST / AMP ASO draft , (2006). [6] EEA, [GMES Initial Operations Land monitoring Services: Annex I - Tender Specifications], European Environment Agency 38 (2011). [7] Härmä, P., Teiniranta, R., Törmä, M., Repo, R., Järvenpää, E., Kallio, E., "The production of Finnish Corine land cover 2000 classification," In International Archives of Photogrammetry and Remote Sensing volume XXXV Part B4, (2004). [8] Haakana, M., Hatunen, S., Härmä, P., Kallio, M., Katila, M., Kiiski, T., Mäkisara, K., Peräsaari, J., Piepponen, H., Repo, R., Teiniranta, R., Tomppo, E., Törmä, M., "Finnish CORINE 2006 project: determining changes in land cover in Finland between 2000 and 2006", Remote Sensing for Environmental Monitoring GIS Applications and Geology VIII Proc. 7110, (2008). [9] Crowley, G., [DMC Data Product Manual], DMC International Imaging Ltd., 127 (2010). [10] EEA, [CLC2006 technical guidelines], European Environment Agency Technical report 17, 70 (2007). [11] Tomppo, E., Halme, M., "Using coarse scale forest variables as ancillary information and weighting of variables in knn estimation: a genetic algorithm approach," Remote Sensing of Environment, 92(1), 1-20 (2004). [12] Tomppo, E., [The Finnish Multisource National Forest Inventory Small Area Estimation and Map Production], In: Kangas & Maltamo (eds.) Forest inventory Methodology and applications Managing Forest Ecosystems Vol 10, Springer, Dordrecht, (2006). [13] Tomppo, E., Olsson, H., Ståhl, G., Nilsson, M., Hagner, O., Katila, M. "Combining national forest inventory field plots and remote sensing data for forest databases." Remote Sensing of Environment, 112(5), (2008). [14] Hatunen, S.,Härmä, P., Kallio, M., Törmä, M., "Classification of natural areas in northern Finland using remote sensing images and ancillary data", Remote Sensing for Environmental Monitoring GIS Applications and Geology VIII Proc. 7110, (2008). [15] Büttner, G., Kosztra, B., [CLC2006 second verification report, Finland], European Environment Agency Mission Report Project 4.1.4: CORINE Land Cover update, 16 (2009).