FOREST STAND CHARACTERISTICS TO BE APPLIED TO MONITORING LINGERING DISASTERS BY SATELLITE REMOTE SENSING AND GIS M. Schardt, H. Kenneweg and H. Sagischewski Technical University of Berlin ABSTRACT The aim of this project is to investigate the applicability of Thematic Mapper data and existing classification methods for forest damage classification of Norway spruce (Picea abies). The test site Harz - one of the most severely damaged parts in Germany - is located between Hannover, Magdeburg and Goettingen. Pure stands of Norway spruce are the most frequent type of silviculture in the Harz Mountains. Today these stands show a variety of forest decline caused by the so-called "Neuartigen Waldschaden" ranking from alomost no injuries in the lower regions to most severe damages like deforestation symptoms in the higher regions. The strongly affected stands are additionally damaged by storm and beetle outbreaks (Ips typographus). From the results of the signature analysis and of former classifications it can be deduced that a distinctive as well as an extensive mapping of forest damage requires the integration of auxiliary information, and for this purpose a GIS is applied including a digital terrain model, forest planning data and soil maps. 1. INTRODUCTION The so-called "Neuartige Waldschaden" first occurred in the early eighties. In contrast to forest damages observed until then (mostly caused by storm and insects), which were very limited in space and time and whose symptoms were well-known, the "Neuartige Waldschaden" occur on a large scale, mainly in the mountainous regions. This forest disease is characterized by discoloration and needle loss and in extreme cases, it leads to the deforestation of relatively large forest areas. These damaged stands are then also more susceptible both to storm damage and to insect outbreaks mostly caused by bark beetles. At first, forest damage inventories were carried out by field sampling methods, later on with color infrared photos (BML 1984 and Hildebrandt 1984). The disadvantage of the former method are that forest decline symptoms due to dead trees are not taken into account and thus do not appear in the damage statistics. This results in a systematic underestimation of forest damages. In contrast to sampling methods, satellite data allow a complete pixel-by- pixel covering assessment on the state of the vegetation integrating also missing trees. For the applicability of these methods it is necessary to define damage classes which correspond both to the requirements of satellite remote sensing and existing field methods of forest damage estimation. The start of the Landsat Thematic Mapper System in 1984 with its improved spatial (30x30 m) and spectral resolution (7 bands: blue, green, red, near infrared, two middle infrared and thermal) offered new possibilities for inventory methods based on Thematic Mapper data. Forster (1989) was one of the first to attempt a classification of extent and intensity of damages in a small investigation area of the Harz Mountains by means of satellite remote sensing. He was able to show that in homogeneous stands forest damages can be classified with an accuracy of up to nearly 70%. Taking Forster's work as a basis, the aim of the joint project between the TU Berlin and the TU Dresden now is to determine the extent of the forest damages of the entire Harz Mountains by means of TM-data for the year 1991, and to record the dynamics of forest damages by means of MSS- and TM-data from 1972 to 1991. In this investigation both the separability of needle loss and deforestation symptoms will be taken into account. The following report deals with the results of the project group of the TU Berlin, whose investigations have been concentrating on the Lower-Saxony part of the Harz Mountains. As support for the satellite classification, a Forest Information System (FIS) was established. The FIS consists of a digital elevation model (DEM) as well as digital forest stand maps and digital soil maps. By merging and overlaying this additional information with the satellite data it is possible to include potentially negative factors such as tree age, changing soil characteristics and illumination conditions in the classification process. Extensive data processing by means of an FIS merely for better classification results would not make proper use of the actual potential of such an FIS and would therefore not justify the amount of work put into it. Thus special attention was given to the possibility of using these data for planning tasks, too. Because of the poor health state of the forest, the forest administration of Lower Saxony plans to convert the widespread spruce monocultures into more stable deciduous and coniferous mixed forest stands or pure deciduous stands (Niedersachsische Landesforsten, 1992). If put into practice, these plans should be accompanied by investigations into how they could be supported by the FIS established during this project. The aim is to show how far satellite classification as a component of the FIS is able to give information on the health state of individual stands, i.e. the urgency of stand conversions; however, the actual spectrum of the statements on further planning which could be supported by such data is much greater. The TU-project described here is part of the European Joint Project for forest damage assessment initiated by the ECE/FAO and UNEP-GRID and entitled "Large Area Operational Experiment for Forest Damage Monitoring Using Satellite Remote Sensing" (Kenneweg et al., 1993). The aim of the project is the development of an integrated classification system and the definiton of integrated and comparable damage classes. Test areas are in Poland, the Slovak Republic, the Czech Republic and in Germany. The investigations are carried out in cooperation with research groups from Finland, France and the USA. 2. DESCRIPTION OF THE TEST SITE AND OF THE DAMAGES The test site of the western Harz is located between Hannover, Magdeburg and Goettingen. The main tree species of the test site are spruce (Picea abies) occurring in the higher regions of the Harz and beech (Fagus sylvatica) growing in the lower parts. Pure spruce stands are the most frequent result of silviculture in the former mining area of the Harz. Today these stands show a variety of forest decline symptoms ranking from almost no injuries to most severe damages. The size of the forested area is about 80,000 ha. The location of the test site Harz is shown in Figure 1. 3. DATA DESCRIPTION AND DATA PREPARATION Satellite Data Thematic Mapper and MSS data of different years (1972-1992) are used for the forest damage inventory and monitoring. By examining MSS data, the reduction of crown density caused by the forest disease will be detected. TM data from 1984-1992 are used to classify changes of needle loss. The satellite images were geometrically corrected into the Gauss-Kruger coordinates using the pass point method. The RMS-error of the geometrical correction amounts to 18-26 meters. Forestry Information System (FIS) From the result of the signature analysis and former classifications it can be deduced that a distinctive mapping of forest disease needs the integration of auxiliary information. Therefore, a Forestry Information System (FIS) was established. The components of the FIS will subsequently be described. A. Digital terrain model (DTM) A digital terrain model is available to eliminate the negative influence of topographical parameters (Schardt, 1987). The topographic parameters, slope, aspect and illumination are calculated from the elevation values provided by the digital terrain model. The digital terrain model is geometrically registered and superimposed on the satellite data. The spatial differences between the DTM and the satellite data amount to about 15 meters due to the y- and x-axis and less than 5 meters due to the z-axis (height above sea level) which is quite satisfactory for classification purposes on a scale of less than 1:50,000. B. Forest management plans In order to integrate stand parameters such as tree age, tree species and stand structure, the forest management plans (70 map sheets consisting of over 15,000 planning units) of the complete Harz Mountains were digitized using the Geographical Information System ARC/Info. Because of the insufficient geometrical accuracy of the forest survey (gaps and overlapping of neighboring map sheets of up to 100 meters) each of the maps had to be geometrically corrected in ARC/Info using a rubber-sheet algorithm and pass points derived from topographical maps. After the geometrical correction, the digitized map sheets are merged together into one forest map containing the complete forest area of the Harz Mountains. As attribute data the digital stand description of the forestry administration could be used. The attribute data containing about 100 different stand parameters are prepared according to the requirements of satellite classifications and computer-aided forest management planning. By combining the attribute data and digitized forest maps, the planning units can be stratified according to single parameters or different combinations of the stand parameters. The result of stratification can be, for example, maps demonstrating the distribution of different age classes, tree species, tree heights or diameter classes. C. Soil maps Digital soil maps are a central component of the Forestry Information System. Together with the digital terrain model, forest management maps, and the forest damage classification these maps will be used for feasibility studies of the Forestry Information System according to forestry management planning. The soil maps were digitized by the Institute of Soil Science of Niedersachsen. The results of this investigation will not be discussed in this paper. D. Topographical maps Digital topographical raster maps were integrated in the Forestry Information System in order to automate the process of geometrical correction of the digitized forest management maps. These maps were made available by the Surveying Administration of Niedersachsen. 4. GROUND TRUTH - SELECTION OF TRAINING AREAS Training areas are test sites used for the signature analyses and for defining the statistical characteristics of signatures of differently damaged stands. The characteristics are used as input for the computer-aided classification. The training areas should be representative for different illuminations, slope and aspect conditions, heights and natural age classes as well as for the different stages of damages occurring in the Harz Mountains. For the selection of training areas, aerial photos were interpreted and field work was carried out. Aerial photos Aerial infrared photos on a scale of 1:6,000 and 1:7,000 were taken for small and representative areas of the Harz Mountains. They are used to estimate the crown density and the percentage distribution of the different damage classes. For the estimation of the crown density, a 1x1-mm grid (6x6 m on the ground) was superimposed on the aerial photos. The relation between grid points covering a tree and those covering a gap between trees can be definend as the crown cover percentage. The percentage of the different needle-loss classes was estimated by interpretation of 3 trees per sample plot on a 30x30-m grid. The interpretation was carried out using the AFL interpretation key (AFL, 1988). Delineation of training areas and verification areas in the satellite image is very difficult without bridging the gap between ground and satellite by aerial photos. Field work Field work was done for altogether 300 training areas to evaluate stand parameters which cannot be derived from aerial photos such as ground vegetation and natural age classes of the stands. 5. SIGNATURE ANALYSIS The statistical analysis of the training areas was carried out by means of the statistical analysis system SAS. The results of the statistical analyses can be summarized as follows: Influence of crown density on the reflection of forest stands From the first results of the signature analysis it can be deduced that even slight differences of crown density have an enormous influence on the reflection of forest stands (Schardt, 1990). The reflection difference can mainly be seen in bands 4 (near infrared) and 5 (middle infrared). The scattergrams in Figure 2 and Figure 3 show the mean values of training areas representing different crown cover percentages separately for different needle-loss classes (x-axis). The y-axis represents the grey values in bands 4 and 5. All these stands belong to the same age and illumination class so that reflection differences are mainly caused by different crown cover percentages of the stands. In band 4 (near infrared) decreasing grey values are associated with a decreasing crown cover percentage within the range from about 65% to 90%. A further reduction of crown density from about 65% to 10% leads to increasing grey values. This phenomenon is due to the increasing influence of high reflecting grass vegetation on the reflection of open stands. In band 5 (middle infrared) a slight reduction of grey values goes hand in hand with a reduction of crown density within the range from about 65% to 90%. Within the range from 65% to 10% a significant increase of grey values is associated with a decreasing crown closure as in band 4. The high deviation of mean values in band 4 and 5 of open stands (10% to 60% crown cover percentage) is due to different types of ground vegetation. Influence of needle-loss on the reflection of forest stands An increasing percentage of needle-loss goes hand in hand with a significant increase in grey values in band 5 (middle infrared) and a slight decrease in grey values in band 4 (near infrared). This signature characteristic can be demonstrated by comparing training areas with different needle-loss symptoms belonging to the same crown cover category. The negative correlation of these two bands due to needle loss symptoms can be shown more clearly be the quotient of bands 4 and 5. Nevertheless, in comparing the signatures of dense stands which are characterized by severe needle-loss symptoms with those of open stands without needle-loss symptoms, no clear separation is possible. From that it can be deduced that the reduction of crown closure and reduction of needles has a similar influence on the reflection of forest stands. Therefore, a satisfactory classification of needle loss without taking deforestation symptoms into account is seen as very problematic. Influence of tree age and illumination on the classification of damage Tree age has a similar influence on the reflection of forest stands as defoliation and deforestation. Independent of the damage class to which they belong, older stands seem to be more damaged than younger stands. Forest stands on south and west slopes show a significantly higher reflection than those growing on north and east slopes. The results of the signature analysis show that a more precise classification of damage is to be expected when integrating digitized forest management plans (tree age) and digital terrain models in the process of classification. 6. DEFINITION OF DAMAGE CLASSES A focal point of this investigation is the definition of a damage class which should correspond both to the requirements of satellite remote sensing and the requirements of the forest administration on forest damage inventories. Remote sensing means measurement of spectral signatures. Spectral signatures of forests are mainly influenced by the amount of needle biomass per pixel. In damaged forest stands this parameter results from both stand density and the average degree of defoliation of the remaining trees. Stand density as a (possible) damage symptom is completely neglected by field sampling methods like ICP-Forest and similar estimation methods, whereas with remote sensing methods it can/must be taken into account. The following matrix illustrates a damage class definition resulting from different stages of defoliation (columns) as well as from different stages of deforestation (rows). This definition was suggested by the working group of the Technical University of Berlin as a basis for discussion. It will be tested under operational conditions. Average crown Defoliation class cover % C0 C1 C2 C3 ------------- -- -- -- -- D 75-88% D0 D1 D2 D2 a 65-74% D1 D2 D2 D3 m 45-64% D2 D2 D3 D3 a 20-44% D3 D3 D3 D4 g 0-19% D4 D4 D4 D4 e Classification of defoliation (columns of matrix) Satellite classifications of forest damage using Thematic Mapper data cannot give any information on the defoliation stages of single trees because from picture elements with a spatial resolution of 30x30 m only an integrated information on approximately 10 to 50 trees can be derived. Because of the heterogeneous spatial distribution of defoliation stages, most of the pixels integrate trees that belong to different defoliation classes. Therefore, it must be examined whether these stages must be assigned to categories according to their respective portion of different defoliation classes or to categories containing a single defoliation class. The latter class definition is only possible if defoliation classes are distributed homogeneously so that mainly one defoliation class occurs within one pixel. In the first phase of the German Harz project the defoliation categories were defined as the portion of strongly damaged trees belonging to the defoliation classes S2, S3, and S4 according to the common damage class definition established by the working group of aerial photo interpretation (AFL, 1988; or VDI, 1990): S0 = 0-10% needle loss S1 = 11-25% needle loss S2 = 26-60% needle loss S3 = 61-90% needle loss S4 = dead tree Because of the nonhomogeneous distribution of damage classes within stands or picture elements of the satellite data, the damage cannot be defined as one single damage class. The classes must rather be defined as the portion of differently damaged trees. So new defoliation categories which better meet the conditions of the spatial resolution of satellite images were defined by Forster (1989): C0 = 0- 10% strongly damaged trees (S2-S4) C1 = 11- 33% strongly damaged trees (S2-S4) C2 = 34- 66% strongly damaged trees (S2-S4) C3 = 67-100% strongly damaged trees (S2-S4) Other definitons classifying the damages of stands are suggested by Schmidtke (1987) and Neumann (1990). One of these damage classifications will be used in this investigation or, if necessary, new damage classes will be developed. Using the damage class definitions described above, the defoliation category of a whole stand can then be calculated from its composition of differently classified pixels. Classification of deforestation / crown density The question of the necessity and practicability of integrating deforestation stages as a damage symptom into the classification is also investigated very intensively in this project. Deforestation due to decreasing crown density can be defined as the percentage of crown cover. The maximal crown cover percentage of a stand without any gap can be assumed as 84-88% (Assman, 1961). Only in very young stands (thickets and young pole timbers) does a crown cover percentage over 88% occur. In the matrix shown above, 5 different crown cover categories are defined. The first category (75-88% crown cover) represents dense stands, whereas the last category (0-19% crown cover) represents very strongly to totally deforested stands. The resulting stand condition categories (D0-D4) are calculated by the total needle loss of a stand, which is the sum of defoliation of single trees as well as the needle loss caused by missing trees. D0 = < 10% needle loss + missing trees D1 = about 11-25% needle loss + missing trees D2 = about 26-50% needle loss + missing trees D3 = about 51-75% needle loss + missing trees D4 = > 75% needle loss + missing trees The practical applicability of this new damage class definition, which takes into consideration the possibilities of damage classification by means of satellite imagery, has to be investigated in this project. Classifications in a smaller area of the Harz Mountains ("Ackerbruchberg area") showed that Thematic Mapper is a suitable instrument to classify the above mentioned deforestation (crown density) classes. Nevertheless, auxiliary data such as digitized forest management plans or other maps providing information about the actual forest distribution are needed for a decision whether or not stand density has to be regarded as a damage symptom, since low stand density may also be the result of silvicultural treatment of forest stands. Without any further interpretation of the classification result, no statement concerning the damage situation of the classified area can be made. Furthermore, detailed knowledge about the influence of the ground vegetation on the reflection of the whole stand is necessary. Advantages of this new class definition - As in former forest damage inventories, the important information of needle loss of single trees in terms of portions of the different damage classes (S2-S4) will be integrated. - The very important information of the stand density which indicates the stability of a stand will also be considered. For a correct interpretation of these new damage classes, knowledge of the classified area must be taken into account or digitized forest maps must be superimposed on the classification results in order to know what state the stand is supposed to be in. - The picture elements of the satellite data integrate trees (more or less damaged) as well as shadow in less deforested stands or ground information in stands with a decreased crown density. This new class definition is therefore more adequate to the signal which is measured by the sensor. It is not possible to isolate the defoliation symptoms from picture elements integrating defoliation as well as deforestation featues. This operation is only possible with the classification of defoliation categories within one deforested stage. Disadvantages of this new class definition - The signal measured by the sensor system does not provide any information about the causes of a low crown density. Here it is pointed out once more that auxiliary information is absolutely necessary to interpret the classification result correctly. Additional information about the silvicultural treatment of the forest can be provided by digital forest maps. For large areas, digital forest maps are not available and cannot be digitized in a reasonable time. Knowledge about the forest is therefore necessary to decide whether the decreased canopy density is caused by silvicultural treatment of the forest or by forest damage. - The problem of recognizing damages caused by air pollution among clearcuts, windthrown areas or something similar cannot be solved if information on regular cutting operations and on disaster (storm, snow, ice, insect attacks, etc.) is not available. - A decreasing canopy density leads to an increasing influence of ground vegetation on the reflectance of forest stands. The typical reflection characteristic of forest stands can be modified strongly by different types and phenological stages of ground vegetation. Thus, a lower classification accuracy of deforestation and defoliation categories can be expected in stands with low densities. Therefore, for stands with very low canopy density, no assessment of different defoliation categories is possible. Detailed field information and knowledge about the reflection characteristics of the forest floor is necessary to estimate the influence of ground vegetation on the classification accuracy. - Grass vegetation of deforested stands sometimes has an identical reflection to the vegetation of agriculturally used areas. Additional information by the categories of forest and nonforest must be added. Another possibility to overcome this difficulty is a multi-temporal approach. 7. CLASSIFICATION The classification was performed separately for different age classes and illumination classes by a simple threshold method using the grey values of band 4 and 5 (near and middle infrared). The stratification of the different age and illumination classes was realized using the digital terrain model and the digitized forest management plans. The comparison of the classification result with aerial photos shows a satisfactory correspondence according to deforestation symptoms. In order to examine the classification result more objectively, an independent aerial photo interpretation for representative areas of the Harz was carried out simultaneously. 8. CONCLUSIONS Satellite data are useful for classifying forest damage when integrating canopy density. Although the establishment of a Forestry Information System is too labor-intensive when used merely for the improvement of classification results, integration of satellite classification into a Forestry Information System can be very useful to support forestry management planning. This information cannot be derived from other already existing maps. Satellite data are important for documenting the temporal and spatial development of forest damage. The ground resolution and the geometrical accuracy of the Thematic Mapper system is not sufficient for overlaying with forestry maps at a scale of 1:10,000 without any further generalization. Therefore, the forest units must be joined together into larger areas in order to minimize the overlay error caused by the edge effect. It is not possible to define one "universal method" for classification of forest damage. Different types of damage, tree species, site conditions and topographical conditions of the forest require different classification strategies. 9. REFERENCES Arbeitsgruppe Forstliche Luftbildinterpreten (AFL). 1988. Auswertung von Color-Infrarot-Luftbildern. Wien und Freiburg. 32 pp. Assman, E. 1961. Waldertragskunde, pp. 99-103. Bayr. Landw. Verlag Munchen. BML. 1984. Waldschadenserhebung 1984, Bundesministerium fur Ernahrung, Landwirtshaft und Forsten. Bonn. Forster, B. 1989. Untersuchung der Verwendbarkeit von Satellitenbilddaten zur Kartierung von Waldschaden. DLR Forschungsbericht (DFVLR-89-06, Oberpfaffenhofen). Hildebrandt, G. 1984. Waldschadensinventur mit Hilfe der Fernerkundung. Z. Flugwissenschaft, Weltraumforschung, 8(1984), Heft5. Kenneweg, H., M. Schardt & H. Sagischewski. 1993. Large area operational experiment for forest damage monitoring in Europe using satellite remote sensing. Proceedings of the International Symposium on Operationalization of Remote Sensing, April 19-23, 1993. Neumann, M. & K. Schadauer. 1990. Waldzustandsinventur, Methodische Uberlegungen und Detailauswertung. FBVA - Berichte, Schriftenreihe der Forstlichen Bundesversuchsanstalt. Niedersachsische Landesforsten. 1992. Waldenwicklung Harz, Fachgutachten. Herausgeber: Niedersachsisches Ministerium fur Ernahrung, Landwirtschaft und Forsten. Schardt, M. 1987. Expositionsabhangiges Reflexionsverhalten van Waldbestanden im Schwarzwald. 2. DFVLR-Statusseminar: "Untersuchung und Kartierung von Waldschaden mit Methoden der Fernerkundung." DFVLR Tagungsband, S. 328-347. Schardt, M. 1990. Verwendbarkeit von Thematic Mapper-Daten zur Klassifizierung von Baumarten und naturlichen Altersklassen. Dissertation an der Forstwissenschaftlichen Fakultat der Universitat Freiburg. Schmidtke, H. 1987. Zur Definition von Schadniveaus fur bestandes- und betreibsweise Waldschadensinventuren. Allgemeine Forstzeitschrift, Heft 37. Verein Deutscher Ingenieure (VDI). 1990. Messen von Vegetationsschaden am naturlichen Standort. Interpretationsschlussel fur die Auswertung von CIR-Luftbildern zur Kronenzustandserfassung von Nadel- und Laubgeholzen. VDI 3793, Blatt 2.