Tuomas Häme, Kaj Andersson and Anssi Lohi; VTT Automation Finland

Herve JeanJean, Philippe Rapaport and Scot Conseil; France

Ian Spence; Swedish Space Corporation

Elisabetta Carfagna; University of Bologna

Michael Köhl and Risto Päivinen; European Forest Institute

Thuy Le Toan; CESBio, France

Shaun Quegan; University of Sheffield, England

Christine Estreguil, Sten Folving and Pamela Kennedy; JRC Space Applications Institute

Abstract

Objective

The objective of the FMERS study was to develop and implement methodologies for the provision of standardised geo-referenced information and statistical information to describe the forests and other wooded land in Europe using optical and microwave space borne remotely sensed data. The target variables were forest and other wooded land as well as proportions of major tree species groupings. The new 10 percent crown cover limit between forest and other wooded land defined by FAO was utilised in this study. In addition to the forest cover classes, which were adjusted to the FAO nomenclature, an extra class ‘temporarily unstocked’ was defined. This class included mainly forest regeneration areas.

The study was divided into two stages: 1) the Pilot Study in which the methodologies were developed and compared at six representative study sites across Europe and; 2) the Regional Mapping that demonstrated forest cover mapping in two large areas in Europe.

The six study sites of the Pilot Study were located as follows: two in the Mediterranean zone, two in the Temperate forest, one in the Alpine region, and one in the Boreal forest. The selection of the sites reflects the increased variability and complexity of forests outside the Boreal forest zone and the uncertainty on the forest resource information in many parts of Central and Southern Europe. On each area, a local expert collected a ground data set.

In the Regional Mapping stage, forest and tree species maps were made from two large regions in Europe, reaching from Southern Italy to Central Finland.

The image data were altogether 65 satellite images. The ground resolution ranged of 10 meters of SPOT to approximately 200 meters of IRS-WiFS.

Most of the optical images, analysed in this study, were from mid to late summer because at that time the growth period was assumed to be in a more stabile development stage than during early summer. Some early-summer images had to be accepted due to clouds. Utilisation of winter images was tested in Finland. The SAR images were from all seasons (Table 1).

The principal ground data, used both for the labelling of the spectral classes and validating the classification results in the Pilot Study, were collected along transects. The transects were parallel or perpendicular strips with a width of 20 metres. Local forestry experts for each site located the transects on the forestry maps or aerial photography, and recorded the values of the target classes along the transects. Using the start and end co-ordinates the transects could be placed on the satellite images, which had been rectified to the same co-ordinate system. The total length of the transects was 1868 kilometres.

Table 1. Image data analyzed in the study.

In the Regional Mapping Stage, no specific ground data were used but the classes were labelled using their spectral reflectance values and all other available data.

Methodology development

The optical image data were calibrated into reflectance values using the 6S atmospheric correction procedure. The calibration made it possible to use the same or very similar parameter values in the image clustering program over all the sites. It also made it possible to construct a ‘reflectance database’ of all the spectral classes defined in this study. All the data were rectified to the same co-ordinate system as the ground transects.

For the optical images the following image interpretation methods were tested:

• an advanced clustering approach;
• the advanced clustering combined with image segmentation;
• maximum likelihood classification;
• regression analysis; and
• combination of the Nilson-Kuusk reflectance model and advanced clustering.

The advanced clustering method has been developed by the co-ordinator of the FMERS study, originally for change detection purposes. It uses the k-means clustering approach with sub-samples from the imagery. These sub samples represent spectrally homogeneous ground targets. After clustering of the samples has been completed, the image is classified pixel by pixel using the statistics that were produced in the clustering process (Figure 1).

The SAR image interpretation methods for forest discrimination are not established. The partners that were responsible for the SAR component developed a new method in this study. The method has three main phases:

• Calibration and geometric co-registration of the multi-temporal image set. The number of images in the data set varied from 2 to 15;
• Multi-temporal filtering. The average temporal change in the backscatter value is computed;
• Spatial filtering to reduce the speckle.

The pixel of the output image is assigned to be forest if the backscattering value is within certain limits, and the seasonal variation of the backscattering is small. The limits could be defined using the ground data but the results suggested that the temporal variation of the backscattering value in forests is consistently less than 2.5 dB.

Two IRS-WiFS image mosaics were compiled from both regions. The location of these regions, the Southwestern region and the Northeastern region, had been defined by the JRC. The image intensities in the ‘direct’ and ‘calibrated’ mosaics were reflectance values. In the so called ‘calibrated mosaic’ an additional relative calibration was done to reduce the effect of reflectance borders between images. These borders were caused by the seasonal differences between the images and also possibly by different atmospheric conditions. The relative calibration was done by selecting a master image in the middle of the mosaic and calibrating the reflectance values of the other images using the reflectance values of mature forests on the overlapping areas of the images.

After preliminary classification experiments the ‘calibrated’ mosaics were selected to the actual classifications.

The phases to produce the calibrated mosaic were as follows:

Sampling system for forest area estimation

The sampling system that was developed in this project aims at producing estimates at pan-European area level for forest and other wooded land using primarily high resolution satellite data. These data have to be combined with ground data to get reliable and unbiased estimates. The statistical estimates can be used as such but also to calibrate satellite maps that have been derived from satellite imagery.

Figure 2. Two stage cluster sampling.

A two-stage area frame cluster sampling was selected as the principal approach (Figure 2). This selection was done using literature and earlier work by the consortium.

From the point of view of the spatial characteristics of the target variables, all of them have a positive spatial auto-correlation. The spatial auto-correlation of the target variables was analyzed by computing stratum-wise correlograms from the CORINE data. The correlograms and the cost function were utilized to compute a homogeneity measure. This measure allows to determine the optimum combination of segment size, number of segments in a cluster and the cluster size.

Results and discussion

The same forest cover classes were targeted both in high and medium resolution image classifications. Altogether 32 classifications of the optical data and 10 classifications of the SAR data were done. The classification performance of the best classifications, using high resolution optical, SAR, and medium resolution optical data, is presented in Figure 3 and Figure 4. The SAR data results are only shown for the Finnish, English, and Polish sites, although the classifications were performed at all sites. This is because in more mountainous areas the algorithm did not give satisfactory results in terms of the transect data validation.

The percentages of correctly classified observations vary from 65 percent for the WiFS classification in England to 94 percent for the SAR classification using fifteen images in Finland. It should be noted, however, that only two classes were discriminated using the SAR data in Finland, i.e. ‘forest’ and ‘other land’. This increases virtually the performance figures of the SAR classifications compared to optical ones, in cases when three target classes were discriminated in the optical classifications.

The Kappa values were sometimes low and illogical which partly reflected problems in the ground data. For instance, the English ground data were almost exclusively from forests i.e. non-forest data were practically missing. In Italy, the generalisation level of the ground data (actually from forestry maps) better corresponded to the resolution of the WiFS data than the resolution of Landsat TM data.

Although radiometric sensitivity of the WiFS instrument appeared not to be high the MSU-SK sensor of the Resurs satellite suffered from more severe radiometric problems (Figure 5). Therefore WiFS was selected as the instrument to the regional mapping.

The performance of the tree species discrimination was on average poorer than for the performance for the cover type classification. This was predictable because the tree species proportions were estimated using a classification procedure although they are continuous variables.

The accuracy of the position of forest cover type borders as derived from satellite data was estimated by accepting different shifts in the location of borders in the ground data and image data. The following alternatives were tested:

It could be concluded that the borderline detection succeeds well using the high resolution data if the borders are clear on the ground (Finland and Poland). The accepted shift of 260 metres was still quite small for the medium resolution data. This shift corresponds to a distance of one millimetre in a 1:250,000 map only. The testing with larger accepted shifts is difficult taken the typical forest patch size in Europe into consideration. Thus, the medium resolution data are too coarse to exactly map the forest borders.

Boreal zone

The classification of forests succeeds well and discrimination of the regeneration areas (temporarily unstocked land) from the agricultural lands is satisfactory. The winter images are not useful. The optimal season is June-August for the optical data. Already the medium resolution image data produces rather accurate forest maps but the poor sensitivity of the presently available instruments at low reflectance levels is still a problem.

Figure 3. Performance of the classification of forest cover as estimated using the percentage of correctly classified observations. The number above the bar shows how many classes have been validated. HR - High Resolution; MR - Medium Resolution.
Figure 4. Performance of the classification of forest cover as estimated using the Kappa coefficient value. The number above the bar shows how many classes have been validated. HR - High Resolution; MR - Medium Resolution.
PLTM1 Advanced clustering with Mask (extract)	A thresholded image of the temporal change image (ERS SAR) of the Polish test site.
PLWiFS1 Advanced clustering (extract)	PLSK1 Advanced clustering (extract)
Figure 5. Classifications from the Polish site. Forested classes as dark tones. The SAR classification result has been median filtered using a 3 x 3 pixel window. The forest/non forest border has been determined using a threshold of 2.5 dB in temporal change. Size of the area approximately 40 km x 40 km).

Temperate zone

The Polish site had rather a clear forest patch structure whereas the forest patches at the English site were small and dispersed. The classification performances in Poland were similar to those in the boreal zone but poorer in England, particularly when the medium resolution data were used. The Polish and English sites may represent the extremes of the European temperate forests in terms of uniformity. If the same level of accuracy in forest mapping is required across the whole zone, a higher resolution satellite data should be used on areas with a dispersed forest structure.

The appropriate resolution a the satellite image can be roughly estimated by comparing the area of three by three pixels of a satellite image to the size of typical forested area. With the IRS WiFS data, the three by three pixel area is 32 hectares, with Landsat MSS 6 hectares, and with Landsat TM 0.8 hectares. If for instance, the typical forested area size is 10 hectares, Landsat MSS would be an appropriate instrument for forest mapping.

Forest classification is difficult with both optical and SAR data and the results are rather poor due to the shadows. In the Regional Mapping stage of FMERS, the classification results in the Alpine zone were better than those in the Pilot Study stage.

The complex landscape of the Mediterranean area produces problems not only in image interpretation, but also in classifying the forests on the ground. The FAO nomenclature with the ten-percent limit for crown cover and five metre length for actual forests causes the border between ‘forest’ and ‘other wooded land’ to be located in the middle of the Mediterranean ‘macchia’ vegetation.

The Mediterranean forests, concentrated in relatively isolated areas, are not as dispersed as the park-type forests in England, for instance. Therefore even medium resolution satellite data may be appropriate for forest mapping at a regional level. Shadows on mountainous areas decrease the classification performance but not as much as in the Alpine zone.

Classification methodology

The principal classification method, advanced clustering, gave better results than any other method at the Finnish and Polish site. It could be concluded that the clustering approach was the best robust method of the tested methods for a European scale forest mapping. An unsupervised approach guarantees that all the variation in the target area can be taken into consideration. A drawback of the k-means method is that it does not take the correlation between the channels into consideration and that the variation of the input channel determines the weight of the channel in the clustering process. These drawbacks can be and were considered by making a reflectance normalisation for each channel during the clustering procedure.

Utilisation of image segmentation usually slightly improved the results, but only with the high resolution data. In the medium resolution data the image resolution is too coarse compared to the forest patch size to make image segmentation meaningful.

The main synergism between the evaluated C-band SAR and optical data could be found useful in the operational completion of the forest mapping exercise. Forest discrimination could be done using SAR, on relatively flat areas that have frequent cloud cover. Such areas exist for instance in the northern latitudes and in the tropics. In addition, the C-band SAR may produce information from very low biomass levels, which is difficult to obtain using the optical instruments.

Figure 6 shows the approximate locations of the classes in the spectral space for the Southwestern mosaic. A final classification, combined from the stratum-wise classifications, is shown in Figure 9. Two parallel strips showing higher reflectance frequencies suggest occurrence of some radiometric problems in the WiFS data.

The stratification clearly improved the results because the ecological variability was better taken into consideration. Not only the variability in the natural conditions but also the radiometric problems of the sensor and inadequate information on the atmospheric conditions in the reflectance computation stage are factors that can increase the effectiveness of the stratification.

Figure 7 and Figure 8 show a visual comparison between WiFS classifications, Landsat TM color composites, and extracts from CORINE Land Cover over two pilot sites of FMERS in Southern and Central Europe.

The north-eastern region mosaic was composed of nine WiFS images acquired during the period 11th August to 5th September 1997. This leads to a variation in meteorological conditions, and a visual inspection of the mosaics showed a significant influence of cloud and haze in many parts of the mosaic.

Sampling system developed

In the simulation for the sampling design development we utilized 8 strata. The sampling rate was approximately 3 per cent but however a minimum number of 3 clusters were allocated also in the smallest strata. The number of selected sampling clusters resulted to be 30 after the intersection of the old CORINE land cover and the ECO-regions.

In most cases, the auto-correlation decreased after the distance of approximately 30 km. In one stratum, representing the Landes and Pyrenees, the auto-correlation remained high up to the distance of 65 km.

The correlograms confirmed that a single stage cluster sampling would be very inefficient, a simple random sampling would be very expensive and a two stage sampling plan gives the highest precision under a fixed budget.

To compute the optimum segment (secondary sampling unit - SSU) size, the following cost function was applied:

C = n (f + h ´ M²);

where n is the total number of segments to be surveyed, M² is the segment size in hectares, M is the side of the square segment, f is the unit cost per segment (independent of the size; it expresses mainly the cost of travel from one segment to the next and of locating the segment on the ground - EUR78 was applied) and h is the cost of surveying one hectare of the segment (EUR8 was applied).

The developed sampling design gives the most precise estimates with the fixed budget. In this design the cluster (primary sampling unit - PSU) sizes and numbers per stratum and segment

(secondary sampling unit - SSU) numbers within a PSU vary. For most strata the optimal size of the PSUs was approximately 50 km by 50 km and the size of the SSUs approximately 30 hectares (Figure 10, Figure 11). The optimal number of the PSUs within a stratum varied from 26 (Northern Temperate forest) to two (Landes and Pyrenees). The number of SSUs within a PSU varied from four to 14.

The results show that the CV% values at a stratum level are high. Thus, for a single stratum the estimates would be poor if the above-described sampling design was followed. The CV% values are particularly high for class 4.2 (predominantly other

wooded area) due to its small area extent. If this variable is of particular interest, a much higher sampling rate should be adopted.

At the level of the area covered by the ‘core’ CORINE", using the optimum size and number (92 in total) of PSUs and the optimum number of SSUs of 30 ha in each stratum (674 in total), we obtain for class 4 ‘wooded area’, an estimate of 1,105,016 km², with a CV% of 6.2 (68732 km²).

The results are strongly influenced by some choices that have been done. If the image prices and processing costs decrease, the number of PSUs per stratum and SSUs within a PSU increase in the optimal design. The total budget for image acquisition and interpretation and for ground truth collection was estimated to be EUR 500,000.

Comparison between WiFS classifications, CORINE , and NUTS databases

The available information sources that were used to validate the WiFS-derived classifications were the CORINE Land Cover database issued by the European Environment Agency and the NUTS (Nomenclature des Unités Territoires Statistiques) statistics issued by EUROSTAT. The FIRS stratification and the forest / non forest map from the AVHRR data (ESA 1992) was also used for comparison.

Only those NUTS II regions whose cloud cover was less than 10 percent in the WiFS classification in the southern mosaic and 25 percent in the northern mosaic were accepted for the comparison. Furthermore, more than 90 percent of the region should be covered by WiFS and CORINE data (75 percent for the northern mosaic in order to select enough regions).

The results of the comparison are:

• The plots between EUROSTAT estimates and WiFS estimates show high correlation, with r˛>0,9 (Figure 12);
• The CORINE Land Cover very closely matched with the NUTS level II statistics. The matching was better and the bias less than with the WiFS classification in the southern region. In the northern region the WiFS classification results looked even better than those from CORINE. Note however that the data-sets for the comparison were limited in the northern region due to clouds and poor CORINE coverage;
• WiFS and CORINE forest statistics were consistent within NUTS II regions;
• The forest area estimates of WiFS classifications were closer to the CORINE estimates than what were the AVHRR forest area estimates of the ESA map.

The WiFS classifications and CORINE LC were also compared at per pixel level but the results were rather poor. The reasons for this may be a combination of local differences in classifications and geometric differences in both data-sets.

Figure 12. Comparison between EUROSTAT forest statistics and the WiFS classification.

Cost benefit analysis

A method developed by the Australian Bureau of Agriculture and Resource Economics (ABARE 1994) was used for the cost analysis. The method was created to analyse the costs of making a multi-annual forest cover map over Victoria, Australia using Landsat TM data.

The full regional forest mapping, including the validation of WiFS-derived forest maps, is estimated to cost EUR 1,233,025 and to last two years. An up-dating was assumed to be done every three to five years. The collection, processing and interpretation of the WiFS scenes represented 30 percent of the total cost, and the validation 70 percent.

Hardcopy Maps

Hardcopy maps were produced for the North-Eastern and South-western regions at a scale of 1:4,000,000 and for the Alpine region of the South-western mosaic at a scale of 1:1,000,000. The maps were reproduced in the CORINE Land Cover projection system (Lambert Azimuthal Equal Area).

Conclusions

The following specific conclusions were made:

Geographic stratification before the clustering improved the results.

The best results are achieved in forest / non forest discrimination, and the most difficult category is the mixed forests. Reliable interpretation of the tree species proportions at a pixel level is unlikely since the pixel size is too large compared to the tree size.

The comparison at the NUTS II level showed a good agreement between the WiFS-derived and CORINE estimates. Comparison between the stratum-wise coefficient of variation values of the estimates from the simulated sampling inventory and the NUTS II level results from the WiFS classifications was surprisingly favorable for the WiFS image classification. The problem with the satellite image classifications is that their performance is difficult to estimate in statistical terms. There are no means to estimate the size of the bias using satellite data only. Availability of reliable forest statistics on one hand and the allowable budget on the other hand determine whether just the existing statistics are utilized to calibrate the results or if new statistical information on target variables is computed through a specific ground sampling.

Visual comparison between the CORINE LC, Landsat TM images, and WiFS classification results showed a good general agreement but also some differences. The differences between CORINE and WiFS classifications were emphasized in the per pixel comparison, which only showed a modest agreement.

There seemed to be general tendencies that the satellite-origin forest maps somewhat underestimate forest cover. However, in the comparisons the WiFS-origin forest area estimates were diminished by the cloud coverage because some clouds were accepted within the comparison areas. It is even possible that the forest area would have been actually overestimated in the northern mosaic if the cloud effect could had been taken into consideration.

Ground survey-aided satellite mapping of forests of the member countries of the European Union was computed to cost EUR 1 233 025 which is approximately three times the budget of the FMERS study. The costs would be EUR 0.13 per square kilometer for a single survey and EUR 0.52/km²/10 years for a continuous procedure in which three surveys are made within a ten-year period. The computations were based on using 35 WiFs scenes which is likely too low a number. Also the other costs may be underestimated.

The clouds were even worse a problem than what had been thought. SAR data could be to some extent used to fill in the cloudy parts of the target area, but using the present C-band SAR systems, the information is restricted to forest / non forest discrimination and on rather flat a terrain. The cloud problems are worst in the mountainous areas where the C-band SAR was not proven to be effective.

The shadowing effect in the optical images due to topography was less harmful than expected.

The new approach in this study, image mosaic compilation before classification worked well. The partner, responsible for the interpretation of the northern mosaic hypothesized that somewhat better results could have been achieved using image by image classification. Such an approach may however be considered as a transient procedure because it is laborious and needs a lot of local ground data. The mosaic-based approach sets somewhat higher requirements to the timing of the image acquisition than the image by image-based approach.

The sampling simulations using the CORINE indicated that the size of the primary sampling unit is often quite large in an optimal design, some 50 km by 50 km. This is because of the high auto-correlation of the target variables. The super high resolution satellite images may not be as an attractive data source for a sampling inventory of European forest areas as were hypothesized because of their small area extent.

The results of this study indicate that the major bottlenecks and unresolved problems are in the interface between the satellite images and other data sources. Also the acquisition of high quality satellite data with resolution of approximately 200 meters seems to be somewhat problematic. The most critical problems may be rather practical than scientific.

The work reported in this paper was carried out under contract from the Space Applications Institute of the Joint Research Centre, Ispra Italy (Joint Research Centre/CEO Contract number: 13105-97-07 F1ED ISP SF). The contract was launched in September 1997 and completed in July 1999. The lead contractor was VTT Automation, Finland, in association with Centre d’Etudes Spatiales de la Biosphere, European Forest Institute, Scot Conseil, Swedish Space Corporation, and University of Bologna. A WWW demonstration of the project can be found at http://www.vtt.fi/aut/rs/proj/fmers/ and also on the EWSE system of the CEO.