Processing of Multitemporal Landsat TM Imagery to Optimize ...

918

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 32, NO. 4 , JULY 1994

Processing of Multitemporal Landsat TM Imagery to Optimize Extraction of Forest Cover Change Features Pol R. Coppin and Marvin E. Bauer

Absfrucf-Digital procedures to optimize the information content of multitemporal Landsat TM data sets for forest cover change detection are described. Imagery from three different years (1984,1986, and 1990) were calibrated to exoatmospheric reflectance to minimize sensor calibration offsets and standardize data acquisition aspects. Geometric rectification was followed by atmospheric normalization and correction routines. The normalization consisted of a statistical regression over time based on spatially well-defined and spectrally stable landscape features spanning the entire reflectance range. Linear correlation coefficients for all bitemporal band pairs ranged from 0.9884 to 0.9998. The correction mechanism used a dark object subtraction technique incorporating published values of water reflectance. The association between digital data and forest cover was maximized and interpretability enhanced by converting band-specific reflectance values into vegetation indexes. Bitemporal vegetation index pairs for each time interval (two, four, and six years) were subjected to two change detection algorithms, standardized differencing and selective principal component analysis. Optimal feature selection was based on statistical divergence measures. Although limited to spectrallyradiometrically defined change classes, results show that the relationship between reflective TM data and forest canopy change is explicit enough to be of operational use in a forest cover change stratification phase prior to a more detailed assessment.

I. INTRODUCTION OST of the world’s vegetation is in a state of permanent flux at a variety of spatial and temporal scales. Hobbs [ 11 differentiates between seasonal vegetation responses, interannual variability, and directional change. The latter may be caused by intrinsic vegetation processes (e.g., succession), land-use and/or other human-induced changes (e.g ., pollution stress), and alterations in global climatic patterns (e.g., global warming). The ability of any system to detect and monitor such fluxes depends not only on its capability to adequately deal with

M

Manuscript received August 12, 1992; revised September 28, 1993 and February 14, 1994. This work was supported in part by the National Aeronautics and Space Administration under Grant NAGW-1431 and by the University of Minnesota under Agricultural Experiment Station Project 42-37. P. R. Coppin was with the Department of Forest Resources, University of Minnesota, St. Paul, MN 55108. He is now with the Department of Forestry and Natural Resources, Purdue University, West Lafayette, IN 47907. M. E. Bauer is with the Department of Forest Resources, University of Minnesota, St. Paul, MN 55108. IEEE Log Number 9402682

the initial static situation (reference data base), but also on its capacity to account for variability at one scale (e.g., seasonal) while interpreting changes at another (e.g., directional). However, not all detectable alterations in vegetative cover are equally important to the natural resource manager. Nor is it guaranteed that all changes of interest will be captured very well or at all by any given change detection system. Of particular relevance to the forest manager are the canopy disturbances caused by short-term natural phenomena, such as insect infestation and flooding, and changes resulting from human activities, e.g., resource exploitation, land clearing for development and reforestation. While the former are likely to be temporary and in many cases self-correcting, evidence of the latter generally remains much longer. To date, satellite-sensor-based monitoring techniques have demonstrated a potential for detecting and identifying areas of certain types of forest cover change on satellite imagery. They have not confirmed, however, the capability to detect changes in land-use/land-cover pattems consistently with acceptable accuracy in a wide range of geographic settings and ecological environments. Jakubauskas [2] appropriately summarizes the documented forest cover monitoring efforts as follows: “most of the change studies have been able to detect particular change aspects over time, but not all have preserved the nature and the spatial variation of these change events.’’ The real problem in digitally deriving accurate cover change information from satellite imagery, however, is representative of the standard remote sensing problem, the maximization of the signal-to-noise ratio. In other words, the contribution of “apparently noisy” pixels to the change image must be minimized with respect to that of the pixels that represent no change or actual forest canopy changes. Causes of such unwanted and/or irrelevant changes can be, for example, differences in atmospheric scattering in the visible bands, differences in water and/ or dust content of the atmosphere at disparate moments in time, temporal variations in the solar zenith and/or azimuth angles, and sensor calibration inconsistencies for separate images. Preprocessing of the satellite images prior to change feature extraction and analysis has as its unique goals the establishment of a more direct linkage between data and biophysical phenomena, the removal of

0196-2892/94$04.00 O 1994 IEEE

COPPIN AND BAUER: EXTRACTION OF FOREST COVER CHANGE FEATURES

919

data acquisition errors and image noise, and the masking binations directly correlating physical vegetative cover of contaminated (e.g., clouds) and/or irrelevant (e.g., ur- characteristics to the transformed features. The use of any two indexes within defined groups would ban land cover) scene fragments. Several researchers have attacked the problem of data provide redundant data. However, the combined results preprocessing from a variety of angles. Duggin and Ro- of the above-mentioned studies remain inconclusive. binove [3] are very explicit in stating that “calibration of Wallace and Campbell [8] aptly indicate that adequate indigital image data to radiance units is absolutely neces- dexes can be found for different purposes and that indexes sary prior to the use of multitemporal or multi-image derived from one analysis may be quite inappropriate in sets.” Markham and Barker [4] provide algorithms for another context. Although 12 major different change detection apcalibrating U .S .-processed Landsat thematic mapper (TM) data to radiance. They further indicate that, for relatively proaches or algorithms have been substantiated in the relclear images, a reduction in between-scene variability can evant literature, few hold a practical attractiveness for the be achieved through a normalization for solar irradiance development of a disturbance monitoring methodology by converting spectral radiance to effective at-satellite re- that ultimately is to become operational in the forest management context (reviewed in [9]). Most methods are flectance or in-band planetary albedo. Atmospheric contamination, or the presence of haze, based on per-pixel classifiers and change information conconstitutes the most severe limitation to the radiometric tained in the spectral domain of the satellite image. Benormalization of satellite data. Several different atmo- cause all digital change detection is affected by spatial, spheric scattering or haze removal techniques have been spectral, thematic, and temporal constraints, the selection proposed [ 5 ] . For remote areas, as is usually the case with of a suitable method or algorithm takes on considerable forest resources, ancillary meteorological and/or atmo- significance. Even in the same environment, different alspheric information is often not readily available. In most gorithms may yield different results. instances, the best that can be done is to perform approxIn addition, it is important to note that monitoring natimate corrections by measuring the return from ground ural resources encompasses a temporal dimension. Vegareas of known reflectance and to use a statistical nor- etative canopies are susceptible to variations over time, malization procedure. including disturbances as well as subsequent recovery. In The literature reveals the use of over 50 different veg- most documented studies, the periodicity of the data acetation indexes for green biomass studies. Use of vege- quisition for short- and mid-term monitoring was detertation indexes not only strengthens the association be- mined according to the availability of satellite imagery of tween spectral data and the biophysical characteristics of acceptable quality. vegetative canopies, but in addition provides a mechanism The lack of any operational satellite-sensor-based manfor data volume reduction. Various authors have studied agement-oriented forest cover monitoring system today the redundancy among these vegetation indexes (e.g., may indicate that it is unrealistic to expect successful change detection of forest canopies via the exclusive use [ti]). Based on their conclusions and incorporating the results of research carried out by many other investigators, of satellite data, particularly TM imagery. The hypothesis three groups can be said to encompass the variability of behind this research, however, proposes that the relationship between forest canopy change and reflective TM data all evaluated vegetation indexes. is explicit enough to be of operational value in a change stratification phase prior to actual aerial and/or ground as1) A group containing, among others, the KauthThomas (KT) soil brightness index, and the first sessment. As a consequence, a TM-based forest cover principal component (PC) of a canonical data trans- change classification can be made reliable enough to eliminate the necessity for further, more detailed monitoring formation. 2) A group with the KT greenness index, the second over areas stratified as “no-change.” In addition, more PC, the normalized difference vegetation index, and specific causal-agent-related assessment activities can then the near-infrared based perpendicular vegetation in- be carried out in “change” areas in which the internal variability has been considerably reduced by the change dex. 3) Another group with the KT wetness index and var- stratification process. The potential for substantial cost savings in inventory and management budgets is unquesious band ratios. tionable, especially where large forested areas are conThe KT indexes involve the application of a technique cerned. of sequential orthogonalization following the GramThe principal goal of this research, therefore, was to Schmidt process [7]. In the case of the TM sensor, the develop an objective methodology for forest change dedata of the six reflective bands are dispersed into a three- tection that would provide a reliable first-level indication dimensional space, defining more precisely two perpen- of forest change for forest managers, objectiveness allowdicular planes (plane of vegetation and plane of soils) that ing for automation and thus the application of the techare called greenness and brightness, a transition zone be- nique over large areas. Although primarily the problem of tween the two, and a third component related to moisture site-specific forest change mapping was addressed, the status. The transformation results in a set of linear com- appropriateness and the limitations of multitemporal TM

920

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 32, NO. 4, JULY 1994

analysis for inventorying nonsite-specific forest changes were also investigated. A secondary objective pertained more to the digital image analysis domain, i.e., the optimal choice of a small number of change features from a large number of possibilities documented in the literature. To reinforce the link to forest management on the ground, the issue of stand versus pixel (where the stand is the spatial management unit on the ground and the pixel the remotely sensed spatial information unit) was dealt with, as well, and the question of an appropriate temporal scale was addressed.

gested in the literature, it was decided early in the experimental design phase to study the dynamics of the forest cover over two-, four-, and six-year intervals. In northern forests, this covers both short and medium-length periods and allows for a subsequent self-consistency (whether the recorded changes over six years are similar to the sum of the recorded changes over two and four years) check. This predetermined temporal resolution was one of the major criteria in selecting the study site, because reference data or the information sources to generate them had to be available.

11. STUDYSITE The study site, 421 km2 in size, is located in the southwestern comer of Beltrami County in north central Minnesota. The geomorphology of the area is northern glacial till plain. The landscape is nearly level to gently rolling with loamy soils that developed on the calcareous till and generally have a high water holding capacity and good drainage. The local climate is continental with wide extremes in temperature from summer to winter. The yearly precipitation ranges from 563 to 640 mm. The area is characterized by a predominant, although not continuous, forest cover encompassing the intricate mixture of tree species and vegetation types that is typical of northem Minnesota. Major tree species include aspen (Populus spp.); birch (Betula spp.); balsam fir (Abies balsamea); jack pine (Pinus banksiana), red pine (Pinus resinosa), white pine (Pinusstrobus) ; black spruce (Picea marianu), white spruce (Picea glauca); tamarack (Larix Zaricina); and northern hardwoods (Fruxinus nigra, Quercus rubra, Tilia spp., Acer spp., and so on). The landscape is extremely fragmented, with an average forest stand area of about 3.4 hectares. The forests in this part of Minnesota may be called complex, with an intricate mixture of different tree species, development stages and stand densities. Aspenlbirch and jack pine are the dominant tree species and together account for about half of the stands. About 60 percent of all stands can be considered merchantable and 50 percent of the forest land area is well stocked. There were about 1500 hectares of overmature forest, mostly aspen, with many of these stands showing symptoms of gradual aspen dieback due to Hypoxylon canker (Hypoxylon mammatum). The forests have been under active management since at least the early 1980’s. Data from a 1982 management inventory were available in the form of field-verified section overlays directly derived from the photointerpretation of 1 : 15 840 scale black and white infrared aerial photography, as original field tally sheets, and as compiled township forest cover maps. Forest plantation records and some data on timber sales and other management interventions, although not complete, existed for the last decade, but were limited to public forest lands. To make the link to midcycle inventory updating possible, and to test the range of temporal resolutions sug-

111. METHODS A. Reference Data An assessment of the usefulness of the Landsat TM sensor for forest cover disturbance monitoring can, under the best scenario, only be as good as the reference data that the processed imagery is evaluated against. It follows that the multidate verification information, commonly but often mistakenly called “ground truth,” must be of a quality guaranteeing thematic as well as positional accuracy. To attain the cartographic precision required to adequately meet the research objectives and to secure information classes that are directly relevant to this study, a multitemporal reference data generation procedure was developed and is described in detail in [9]. The procedure incorporated various steps: the creation of an accurate and stable map base, information extraction and data transfer from the aerial photography, validation via original 1982 inventory tally sheets, plantation records, timber sale documents, and management data from the local field foresters, and digitization in raster format with grid size equal to the TM pixel dimensions (100 percent coverage). Multitemporal 35-mm vertical aerial photography with an average nominal scale of 1 : 100 000 existed for the entire experiment site. The set consisted of a mixture of color and color infrared (CIR) slides for May 1984, CIR slides for May 1986, and CIR slides for early October 1990. Because the reference data also needed to serve other purposes, all possible information was extracted from the aerial photography with respect to the forest cover condition at the individual points in time (1984, 1986, and 1990), as well as to the dynamics of the cover canopy over the time intervals of interest (two, four, and six years). For more details see [9]. It is appropriate to stress that only the availability of detailed ancillary information in hardcopy format (1982 field inventory reports, township forest type maps from different years, plantation and timber sale records, meteorological data, and so forth) and/or in the form of professional knowledge accumulated by the local forest managers made it feasible to organize the observed forest cover changes in a causalagent-driven thematic framework such as shown in the first column of Table I. However, this classification could not be made totally exhaustive, nor mutually exclusive, nor fully consistent. For example, although on-site observations in the summer

~

COPPIN AND BAUER: EXTRACTION OF FOREST COVER CHANGE FEATURES

TABLE I FORESTCANOPYCHANGE EVENTS

B. Imagery

Aruial Photography and Ancillary Dam

Aerial Phot-

charm clwrcut + natural reganeration charcm + plsntation brush cleariw + plantation floodiw hishwdiw. wlective cut,dieback w~~ veg.ution "oval storm &maps ldter 4 and 6 years1

diwllpamnce of canopy diwpp.annce of Campy with

storm damass (after 2 veard

spatially fragmented canopy

recent sorm damass

wdv " w & n devdoyww plantation development

higher density of campy

net canopy gain or increnmnt

no change

stable campy

no chanw

rows of

Only

92 I

sadlite Inqpry net campy loss or depletion

&MS

lower d"ity Of unow p o c h r ( l d c.mpy

and fall of 1989 and 1990 confirmed the periodic incidence of large-scale insect infestations in young aspen populations (leaf roll), no such phenomena could be reconstructed from the available data sources. Also, selective cutting or localized clear-cutting was going on in forest stands classified as storm-damaged. It must be noted also that young tree populations that had not yet attained full canopy closure at the initiation of a monitoring interval were classified as regeneration or plantation development, regardless of their subsequent development pattem. To be successful, any digital change classification scheme needs to be totally exhaustive, mutually exclusive, relevant to the resource management objectives, and consistent over time. A regrouping of the first canopy change event classification based on photo parameters only (ancillary information sources eliminated) resulted in seven categories (column 2, Table I) that met the criteria of exhaustiveness and consistency much better, but still failed with respect to the principle of mutual exclusiveness. Moreover, all attempts to generate digital spectral change signatures according to either 12 or seven change events proved unsatisfactory with major confusion remaining among the classes within the four basic canopy change groups (column 3, Table I). The change detection methodology proposed here is therefore built on four basic change categories that meet the predefined operational and scientific criteria: no change (closed canopy remained closed, or canopy lacking), canopy depletion (net overall canopy loss), recent storm damage (structural/textural canopy changes), and canopy increment (process of canopy closure). In the final data base all nonpublic lands were masked out. A total of 3715 cover type units were mapped in 1984 and subsequently monitored for change. Over the entire study area and for the two-year period, the reference data identified 259 hectares of forest cover as being affected by canopy depletion, 439 hectares by canopy increment, and 165 hectares by storm damage. For the four- and sixyear periods, the numbers are, respectively, 736 and 1229 hectares for canopy depletion and 838 and 817 hectares for canopy increment. At the stand level, 229 stands were classified as changed over the two-year interval, 465 over the four-year interval, and 759 over the six-year interval.

The multidate TM imagery necessary to support this research was obtained in conjunction with a NASA-sponsored ''Satellite Inventory of Minnesota Forest Resources'' project (NRA-87-166). The calendar acquisition dates for the TM data were chosen not on a specific anniversary day but within an anniversary window. The July-August window was selected following these criteforest cover phenological stability; lowest monthly percent cloud cover, based on the examination of the Landsat imagery available for the area of interest since 1980; lowest seasonal soil moisture content; and minimal sun angle effects (minimal variation in solar zenith angles between dates in this window). The acquisition window was centered a bit later than the date the forest vegetation is seasonally fully mature. This provided flexibility because it permitted the actual acquisition dates (August 13, 1984, July 18, 1986, and July 13, 1990) to be chosen from satellite overpasses either somewhat later or earlier than that of the true anniversary date, and in the driest, most cloud-free period of the year.

C. Data Optimization Fig. 1 summarizes, in chronological order of execution, the preprocessing routines the TM data were subjected to in order to maximize their relationship with green vegetation features and standardize their information content over time. The thermal TM band (TM6) was excluded from the methodology because other investigators have shown that, for identification of surface types, thermal information is not readily associated with that in the reflective part of the spectrum, which in tum may lead to misinterpretation and/or spurious classification [ 101. Moreover, forest cover seems to exhibit little variation in the 8-12 pm region [ 111. Basically, four procedures were involved: data calibration, scene rectification and registration, atmospheric normalization and correction, and interpretability enhancement via the generation of vegetation indexes. As a first step, the raw digital numbers of the six reflective TM bands (TM1- TM5 and TM7) were calibrated to effective at-satellite or exoatmospheric reflectance via the algorithms proposed in [12]. The calibration takes on even more significance because ratio-based vegetation indexes were computed at a later stage. Spectral radiance extremes for the TM sensors were determined from the TM postcalibration dynamic ranges of U. S.-processed data. Mean solar exoatmospheric irradiances at the particular dates were extracted from the TM scene header records. To expand the obtained multidate minimum and maximum at-satellite reflectance values to an integer 8-bit range (0-255), the data were scaled (multiplication factor of 3.07) and rounded to the nearest full digit, effectively

922

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 32, NO. 4, JULY 1994

Atmospheric Normalization and Correction

Interpretability Enhancement Vegetation Index Generation

(

Vegetation Index Images

Fig. 1 . Preprocessing routines for data optimization.

putting them on a reflectance scale which runs from 0 to 83 percent. Because the terrain relief is negligible over the experiment site (no discemable slopes, relief gradient of 72 m over a 46.6 km north-south transect) the TM scenes were individually rectified and registered to the appropriate UTM grid using a common set of ground control points, a second-order polynomial warp function, and nearestneighbor resampling. The residual mean square error for the GCP’s that were used to calibrate the warp function was 0.23 pixels. This met the condition that, whatever resampling technique is applied, it must always result in the rectified image being within to pixel of the reference map, especially if differencing or ratioing change detection is to be performed [ 131. The atmospheric correction procedure applied combined two major components: atmospheric normalization and transformation to ground reflectance. A statistical regression approach was used to normalize the satellite data over time. Preference was given to a technique based on spatially well-defined spectrally and radiometrically stable features [14] over a method that makes use of ra-

a

firm that atmospheric conditions were uniform over the subscene at each of the three points in time. As such, the procedure does not eliminate the variability across the

in atmospheric composition and optical thickness, but it standardizes the conditions over the subscene quite effectively. The five target surfaces were known not to have been- the object o f a n y disturbance between 1984 and 1990, were easily identified in the subscene, and together covered about the entire range of reflectance values in the multidate imagery. The five- features were a clear, deep, oligotrophic lake; a dense, mature, even-aged, homogeneous red pine stand; a large flat asphalt roof; an undisturbed gravel-covered area; and a concrete aircraft parking slab. Nearly constant illumination and viewing geometry produced consistent reflectances from one date to the next. Both the 1984 and the 1990 data sets were standardized to the conditions of the 1986 image. The assumption that a linear relationship exists between pairs of individual band pixels over time [16] was tested against a secondorder polynomial model fit and against a logarithmic model fit, and clearly proven correct. As a consequence, the between-image residual error due to atmospheric transmission and path radiance variations over time was minimized over the subscene by applying a simple linear regression model of the form (reflectancehel)A = a

+b x

(refle~tance,~,~)~.

The unequivocal linear character of the relationship between band-specific target reflectances over time is illustrated in Figs. 2 and 3 and the model coefficients, t values of the slope coefficients, and R2 values are given in Table 11. The slope coefficients of the linear model fits were always significant at a = 0.05. Once the scaled reflectance factor data were normalized over the temporal dimension, a simple dark-object subtraction technique for atmospheric scattering correction was applied to all dates separately. The ground reflec-

923

COPPIN AND BAUER: EXTRACTION OF FOREST COVER CHANGE FEATURES

-3.183123

8o

+

1.017713(X)

Rz = 0.9875 Y

t

P

9

TABLE I1 LINEAR MODELFITTING FOR ATMOSPHERIC NORMALIZATION TM band

lnterwpf

t- Value of Slope Coefficient

R'

Regression of 1986 data on 1984 data

-5.271798 -2.819122 -2.139920 -3.183123 -0.608076 -1.642934

f

slow Coefficient

1.068317 1.042580 1.054408 1.017713 0.901965 0.879725

11.28 26.05 21.98 15.42 35.13 24.74

0.9770 0.9956 0.9938 0.9875 0.9976 0.9951

Regression of 1986 data on 1990 data

-10.77912 -4.563112 -3.279962 -8.955492 0.419620 -1.420203

1,280274 1.1 21398 1 .lo8905 1.101304 0.928393 0.892908

15.46 16.87 15.23 26.98 20.06 81.95

0.9876 0.9896 0.9872 0.9959 0.9926 0.9996

The critical t value for statistical significance of the slope coefficient is 3.18 (with three degrees of freedom and a = 0.05).

0' 0

'

20

40

60

80

100

TM4 1984 (scaled %) Fig. 2. Linear model fit between 1984 and 1986 calibrated TM data (scaled reflectances).

Y = -8.955492 R2 = 0.9959

+

1.101304(X)

8ol u

60

Qo

due to scattering is inversely proportional to wavelength, the bias is greatest for TM band 1 and least for TM band 4. The mid-infrared bands were assumed to be unaffected by the phenomenon. No attempt was made to compensate for atmospheric absorption effects, as the ancillary data required for this purpose were not available. While it is commonly accepted that the six reflective TM bands fit in only three dimensions of spectral feature space, no consensus can be found in the literature on which indexes or data transformations represent these features the best in the context of green vegetation assessment. It was therefore decided to investigate those seven indexes (VI'S) that were deemed most promising by other investigators (Table 111). Further analysis would then determine which three of those indexes carried the most relevant information for subsequent successful change detection over forest cover. The tasseled cap transformation coefficients for reflectance factors were taken from [ 181 and the normalized difference vegetation index was computed via a computationally more efficient, but functionally and linearly equivalent formula [ 191.

40

0

20

40

60

a0

100

TM4 1990 (scaled %) Fig. 3. Linear model fit between 1990 and 1986 calibrated TM data (scaled reflectances).

tance of a clear deep lake (as the dark object) for the first four TM bands was derived from field measurements of the spectral characteristics of natural waters [17] and scaled to the same dynamic range (8 bit) via the alreadydefined multiplication factor. Because the scene radiance

D. Change Feature Extraction The processing steps and the subsequent verification phase are illustrated in flow chart format in Fig. 4. Change feature extraction was approached simultaneously from two different angles. First, an image differencing algorithm was applied. It was modified to a standardized version to minimize confusion among change values that are numerically equal, but are describing different change events: change indicator (VI time 1 - VI tlme,)/(VItime

1

+ VI time,).

Second, the bitemporal VI pairs were subjected to a selective multitemporal linear transformation routine, where the resulting second principal component (PC,) is assumed to represent the change of the VI over the time interval between the data acquisitions. change indicator = PC,(VI t,me,,VI

IEEE TRANSACTIOINS ON GEOSCIENCE AND REMOTE SENSING, VOL. 32, NO. 4, JULY 1994

924

TABLE IV FOR EACHOF THE TIMEINTERVALS (TWO, CHANGE FEATURES GENERATED FOUR,AND SIXYEARS)

Vegetation Index Bitemporal Pairs

I Code

Description

i Change

Computation

1

2 3 4

5 6

Reference Data

7

8

Subsetting and Masking

9 10 11 12

13 14

Standardized difference of brightnesses Standardized difference of greennesses Standardized difference of wetnesses Standardized difference of Crippen’s NDVl’s Standardized difference of green ratios Standardized difference of MIR ratios Standardized difference of red bands Second principal component of brightnesses Second principal component of greennesses Second principal component of wetnesses Second principal component of Crippen’s NDVl’s Second principal component of green ratios Second principal component of MIR ratios Second principal component of red bands

Change Classification

(

Change Images

Accuracy

)

Assessment

Analysis Fig. 4. Change feature extraction and accuracy assessment routines. TABLE 111 VEGETATION INDEXESAND THEIRREPORTED INFORMATION CONTENT Index

Brightness

Greenness

Wetness

NDVl ICrippen)

Formula

0.2043~TMlI+0.41581TM2)t 0.5524(TM3) +0.5741(TM4) + 0.3124(TM51+0.2303(TM71 -0.16031TM11-0.2819(TM2l0.49341TM3) 0.79401TM4b 0.00021TM5)-0.14461TM71 0.031 51TM1) +0.2021(TM2) 0.3102(TM3) +0.1594(TM41. 0.6806~TM51-0.61091TM7I TM41lTM4 + TM3)

-

Green ratio

TM41TM2

MIR ratio

TM41TM5

Red band

TM3

Reponed Detection Capability

Source

Soil characteristics

1221

Green canopy characteristics State of vegetative cover

1231 161

Soil moisture status 1221 Full canopy closure achievement 1221 Vegetation/soil relative mixture 1221 Major canopy alterations 1241 Coniferous biomass 1241 Canopy closure 1251 Insect defoliation 1211 Wetland characteristics 1261 Water body identification 1261 Flooding assessment 1261 1271 Percent foliar loss Forest decline 1281 Clear cutting 1201.1291.1301

All-in-all, 14 change features were generated for each of the three time intervals of interest (two, four, and six years). They were believed to consolidate among them the maximum association between satellite sensor signal and the object of interest: changes in the forest canopy. They are summarized in Table IV.

Except for the traditional thresholding techniques, very few other approaches on change information extraction from multitemporal data have appeared in the literature. To fully exploit the bitemporal change information content of the generated change feature data sets, multivariate statistical analysis concepts were used to specify a set of discriminant functions dividing the spectral measurement space into appropriate decision regions. A pattern recognition methodology, based upon the characterization of the change classes of interest via the analysis of representative data subsets, commonly called training signatures, and the use of the latter to classify the entire data set by means of the maximum-likelihood (M-L) decision rule, was selected for change information extraction. The M-L rule assumes normal probability density functions. This assumption was confirmed by a visual inspection of the change feature histograms. A two-pronged approach for signature generation was chosen. The spatial location (in row/column numbers) of “seeds” or center pixels for the different change events was visually extracted from the rasterized reference data, and the spectral change polygons that define the areal extent of the signature in the change features were allowed to “grow” from these seeds following spectral-distance decision rules. Only those signatures that matched the spatial outline of the change polygons on the reference change maps in a conservative way (smaller or equal areal extent and form) were retained. This resulted in 31, 39, and 40 change event signatures for the two-, four-, and six-year time intervals. There was no doubt that the data redundancy, prevalent among the seven vegetation indexes, carried over into the change features. It was therefore of primary importance to choose a subset of change features while maintaining the maximum statistical separability among the change signatures. The Jeffries-Matusita (J-M) distance for best minimum separability, which takes into account the covariances as well as the mean vectors of the change sig-

~

COPPIN AND BAUER: EXTRACTION OF FOREST COVER CHANGE FEATURES

TABLE V CHANGE FEATURE SELECTION V I A J-M DISTANCE COMPUTATIONS TwD-Year Change Interval Best Minimum Separability

Best Average Separability X feat.

av. JM dist.

2 3

1341 1369 1382 1391 1396

4 5 6

mi". JM dist.

448 874 974 1031 1049

features 1.9 1.9.12 1.8.9.12 1 1.2.8.9.12 1,2,5,8,9.12

av. JM dist.

1315 1366 1382 1391 1396

min. JM dist.

695 895 974 1031 1057

features 1.12 1.2.12 1.8.9.12 1.2.8.9.12 1.2.8.9.12.13

Four-Year Change Interval Best Average Separability

I fast.

2 3 4 5 6

a". JM dist. min. JM dist.

1302 1356 1377 1387 1393

395 758 779 824 883

features 8.13 9.10.13 3;9.10.12 1.9.12.13.14 1.8.9.12.13.14

Best Minimum Separability av. JM die.

min. JM dist.

1270 1356 ... 1353 1366 1386

690 775 871 945 996

features 9.13

-. -. -

8.10.13

2.6.9.12 2.4.7.8.10 3.4.7.8.10.14

Six-Vear Change lntsrvrl Best Minir"

BeR Average Sepanbilii X feat. 2 3 4 5 6

av. JM dist. min. JM dist.

1286 1308 1321 1329 1339

227 277

304 327 323

features 1.2 1.2.8 1.2.8.9 1.6.8.9.13 1.2.7.6.9.14

a". JM dist.

1283 1307 1315 1329 1333

Separability

mi". JM dist. 247 294 323 357 376

features

1.9 1.8.9 5.8.9.12 1.5A9.12 2,4,5,9,10,12

Best separability among signatures is attained when J-M distances

=

1414.

natures, was used as a measure of statistical divergence to select the best change feature data sets. Selected features are outlined in Table V. Because the statistical significance of the extra separability gained by adding features that lay in the saturation zone of J-M distance versus the probability of correct classification is doubtful, combinations of more than six change features were eliminated from consideration. To diminish the salt-and-pepper appearance that is so common for pixel-based digital classification results, and to approximate somewhat the stand concept, a 4 X 4 mode filter was run through the change images. The size of the filter was determined from comparative classification trials with highest average class accuracy as the deciding factor. Finally, complete enumeration and a discrete multivariate statistical analysis technique (Kappa coefficient) were utilized to determine the change detection accuracy. The Kappa statistic removes the element of chance agreement and is computed from the classification error matrix [31]. The larger its value, the more accurate the classification results. IV. RESULTSAND DISCUSSION Spatially accurate forest cover monitoring requires the precise registration of the multidate imagery. It was found that, with standard Landsat TM imagery, a "within-pixel shift," or residual locational error below pixel dimension, could not be corrected for. Apart from rendering the selection of ground control points for scene rectification more difficult, the phenomenon also somewhat degrades the areal assessment of the change events, specifically at the changeho-change boundary. The operational user will

925

have to accept this within-pixel shift as a limitation inherent to any digital-imagery-based change detection methodology. It is evident from Table VI that a statistical accuracy assessment of the change detection outcome results in Kappa coefficients that are slightly higher for the fourand six-year time intervals than for the two-year interval (the coefficients are significantly different at the a = 0.05 level). It may be inferred that the optimal time span for management-oriented forest cover monitoring in northern temperate forests is towards the upper end of the six-year range. The digital methodology proved furthermore highly consistent. The only exception occurred for aspen regeneration. For example, a clear-cut (canopy depletion) over the two-year period, followed by the establishment of natural regeneration in the form of root suckers over the following four-year period was in many cases not detectable over six years. This is because a successfully established young aspen population with its closed and rather light canopy is very difficult to differentiate on spectral-radiometric grounds from the generally more open mother stand with its usual admixture of other species and/or shrubby understory. No single one-dimensional vegetation index or change feature can be expected to summarize the information in multidimensional spectral data space. Data redundancy in the number of vegetation indices and change features was built into the research design in order to be able to subsequently identify those change features carrying the most relevant information. Although the inclusion of six change features results in the largest average J-M distance (Table V), a reduction to five or even four features decreases the Kappa values of the change detection process only slightly (e.g., for the six-year interval to 0.81 and 0.79). The five most prominent change features (in descending order of number of occurrences in Table V) are the standardized difference of brightness, the second principal component of greenness, the second principal component of brightness, the second principal component of the green ratio, and the standardized difference of greenness. This points to the Kauth-Thomas brightness and greenness indexes and the green ratio as the vegetation indexes with the most relevant forest cover change information for digital monitoring. By design, this forest cover monitoring methodology was limited to the four change events enumerated in column 3 of Table I. A digital TM-based evaluation and ranking of forest cover change events beyond that which is feasible with only spectral-radiometric information would require the incorporation in the GIS framework of extensive and often complicated artificial intelligence capabilities. Their function would be to parallel the interpretative procedures employed by the photointerpreter during reference data generation. Although this is the subject of ongoing research, at the time it was beyond the scope of this study. Also under investigation, but not considered in the investigation discussed here is the relationship between quantitatively assessed canopy change

926

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 32, NO. 4, JULY 1994

TABLE VI ACCURACY ASSESSMENT RESULTS Period

‘84-’86 ‘86-90 ‘84’90

Overall Accuracy

97.0% 95.7% 93.6%

Average Class Accuracy

78.7% 90.0% 91.2%

Kappa Coefficient

0.76 0.83 0.82

events (alterations in biomass, leaf area index, and so on) and their optical effects. This might be another avenue to connect different causal agents and intensities of change phenomena to different spectral-radiometric effects, on the condition that a definite link can be authenticated between the origin of the change and its optical impact on the forest canopy. Thematic accuracy assessment was performed on an independent data set. At the pixel level, the numbers summarized in Table VI without doubt indicate operationally acceptable success rates. A merging of the thematically related and at first impression nonexclusive “canopy depletion” and “storm damage” classes for the two-year period conveyed only a minor accuracy gain: Kappa coefficients of 0.78 versus 0.76 (the difference is statistically significant at the cr = 0.05 level) and average class accuracies of 84.8 percent versus 78.7 percent. The application of a mode filter with its visual homogenizing effect represented a crude attempt to bring the pixel-based classification results closer to the stand concept as embodied in the reference data. Overall, change detection at the stand level also gave origin to operationally acceptable results, i.e., 94 percent accuracy for the six-year monitoring period. A detailed examination led to the following inferences that remain valid irrespective of the character of the change event. 1) All stands smaller than 1 hectare that had disturbances reported in the reference data and were totally surrounded by undisturbed forest cover were lost as changes during the change detection process. 2) All stands smaller than 1 hectare that belonged to the no-change population and were either completely encircled or largely surrounded by disturbed stands were misclassified and clumped together with the nearest change event. 3) Some undisturbed larger stands, ranging in size from 1 to 2.6 hectares, that had a distinct elongated form (strip) and were wedged in between changed forest cover, were also misclassified in that a majority of their pixels were shown as changed. Over the six-year period, 759 stands out of 2306 were reported in the reference data as having been affected by change events. The digital change detection missed only 45 completely (all pixels that made up the stand classified as no-change). Stand size was the limiting factor, because 43 of these stands were smaller than 1 hectare. In the other 714 stands, change was detected in at least 50 percent (and very often in more than 90 percent) of the total number of pixels that made up the individual stands.

V. CONCLUSIONS A reliable, objective, digital methodology for forest change detection with Landsat TM imagery was developed and applied successfully in midlatitude forests. The fact that the technique as presented here has been incorporated in the “Annual Forest Inventory System” (AFIS) proposed at the end of 1992 by the U.S.D.A. Forest Service, North Central Forest Experiment Station, in cooperation with the Minnesota Department of Natural Resources, Division of Forestry, testifies to the operational viability of the endeavor in its present form [32]. If and when a higher level of precision is sought in the disturbance information, the incorporation of expert system technology may simulate the interpretative capabilities of the human operator, but more research is needed. From the digital image analysis perspective, it was found that changes in brightness and greenness identified the most important forest canopy change features, and that these can be adequately expressed either as normalized differences or second principal components. It must be noted, however, that postclassification filtering, applied to more closely approach the stand entity, eliminated small but real features of interest. Corollaries of this research have also demonstrated that, in contrast to what universally happens in photo interpretation procedures, a digital sensor such as the Thematic Mapper does not collect information at the stand level (stand being the forest management unit). The TM’s assessment unit is the pixel, and only insofar as the stand is spectrally-radiometrically homogeneous, will clusters of pixels aggregate into polygons that approximate the areal extent of stands as individualized in traditional forest cover maps. Whereas the loss of the stand concept from the digital evaluation process can represent a disadvantage, the fact that information is provided at the substand level somewhat counterbalances this drawback. From the forest resource management perspective, it has been substantiated that the system is capable of adequately detecting negative forest canopy perturbations, although this occurred at the pixel level and not necessarily at the stand level. Moreover, wind damage can be differentiated from other canopy depletion events when the storm has transpired the year before the disturbance assessment. Conifer plantation development (or in the case of jack pine, natural regeneration development) can be very satisfactorily monitored before full canopy closure is attained. A misclassification into “canopy depletion,’’ or even the absence of the “canopy increment” phenomenon, then often highlights local deficiencies within the developing stand. Aspen regeneration monitoring remains more problematic after the first two to three years. REFERENCES [l] R. J. Hobbs, “Remote sensing of spatial and temporal dynamics of vegetation,” in Remote Sensing of Biosphere Functioning, R. J . Hobbs and H. A. Mooney, Eds. New York: Springer-Verlag, 1990, pp. 203-219. [2] M. E. Jakubauskas, “Utilizing a geographic information system for

COPPIN AND BAUER: EXTRACTION OF FOREST COVER CHANGE FEATURES

[3] [4] [5] 161

[7]

[8] [9]

[lo] [ll] [12] [13] [14] [15]

[16] [17]

[IS] [19] [20]

[21] [22] [23]

[24]

[25]

vegetation change detection,” in Tech. Papers 1989 ASPRS/ACSM Conv., vol. 4 , Baltimore, MD, 1989, pp. 56-64. M. J. Duggin and C. J. Robinove, “Assumptions implicit in remote sensing data acquisition and analysis,” lnt. J. Remote Sens., vol. 1 1 , 1990, pp. 1669-1694. B. L. Markham and J. L. Barker, “Radiometric properties of U.S. processed Landsat MSS data,” Remote Sens. Environ. vol. 22, pp. 39-71, 1987. Y. J. Kaufman, “The atmospheric effect on remote sensing and its corrections,” in Theory and Applications of Optical Remote Sensing, G . Asrar, Ed. New York: Wiley, 1989, pp. 336-428. B. L. Soliday, “Identification of data redundancies in vegetation index models: An application of Landsat spectral reflectance digital data,” in Proc. IGARSS’89, Vancouver, B.C., Canada, 1989, pp. 1339- 1342. R. J. Kauth and G. S. Thomas, “The tasseled cap-A graphic description of the spectral-temporal development of agricultural crops as seen by Landsat,” in Proc. Symp. Machine Process. Remotely Sensed Data, Purdue Univ., West Lafayette, IN, 1976, pp. 4b:414b:51. J. F. Wallace and H. Campbell, “Analysis of remotely sensed data,” in Remote Sensing of Biosphere Functioning, R. J. Hobbs and H. A. Mooney, Eds. New York: Springer-Verlag, 1989, pp. 297-304. P. R. Coppin, “The change component in multitemporal Landsat TM images: Its potential for forest inventory and management,” Ph.D. dissertation, Univ. Minnesota, St. Paul, MN, 1991. J. C. Price, “The contribution of thermal data in Landsat multispectral classification,” Photogrum. Eng. Remote Sens., vol. 47, 1981, pp. 229-236. T. J. Schmugge, F. Becker, and Z.-L. Li, “Spectral emissivity variations observed in airborne surface temperature measurements,” Remote Sens. Environ., vol. 35, 1991, pp. 95-104. B. L. Markham and J. L. Barker, “Landsat MSS and TM post-calibration dynamic ranges, exoatmospheric reflectances and at-satellite temperatures,” EOSAT Landsat Tech. Notes, no. 1 , 1986, pp. 3-5. J . R. Jensen, “Urban change detection mapping using Landsat digital data,” Amer. Cartograph., vol. 8 , 1981, pp. 127-147. V. Caselles and M. J. Lopez Garcia, “An alternative simple approach to estimate atmospheric correction in multitemporal studies,” Int. J. Remote Sens., vol. 10, 1989, pp. 1127-1134. F. G. Hall, D. E. Strebel, J. E. Nickeson, and S. J. Goetz, “Radiometric rectification: Toward a common radiometric response among multi-date, multi-sensor images,” Remote Sens. Environ., vol. 35, 1990, pp. 11-27. A. J. Richardson, “Relating Landsat digital count values to ground reflectance in optically thin atmospheric conditions,” Appl. Opt., vol. 21, 1982, pp. 1457-1464. D. E. Bowker, R. E. Davis, D. L. Myrick, K. Stacy, and W. T. Jones, “Spectral reflectances of natural targets for use in remote sensing studies,” NASA Ref. Pub. No. 1139, Washington, D.C., 1985. E. P. Crist, “A TM tasseled cap equivalent transformation for reflectance factor data,” Remote Sens. Environ., vol. 17, 1985, pp. 301-306. R. E. Crippen, “Calculating the vegetation index faster,” Remote Sens. Environ. vol. 34, 1990, pp. 71-73. T. H. Hame, “Satellite image aided change detection,” in Remote Sensing-Aided Forest Inventory. Helsinki, Finland Dep. Forest Mensurat. Manage.: Univ. Helsinki, 1986, Res. notes no. 19, pp. 47-60. R. F. Nelson, “Detecting forest canopy change due to insect activity using Landsat MSS,” Photogram. Eng. Remote Sens., vol. 49, 1983, pp. 1303-1314. E. P. Crist and R. C. Cicone, “Application of the tasseled cap concept to simulated thematic mapper data,” Photogram. Eng. Remote Sens., vol. 50, pp. 343-352, 1984. D. P. Rice, W. A. Malila, and R. F. Nalepka, “The threshold of detection of vegetative canopies using remotely sensed data,” USDA Forest Serv. Final Rep. no. 124000-5-P, Environment. Res. Instit. Michigan, Ann Arbor, MI, 1979. S. A. Sader, “Remote sensing investigations of forest biomass and change detection in topical regions,” in Satellite Imageries for Forest Inventory and Monitoring; Experiences, Methods, Perspectives. Helsinki, Finland: Dep. Forest Mensurat. Manage., Univ. Helsinki, 1988, Res. notes no. 21, pp. 31-42. R. F. Nelson, “Detecting forest canopy change due to insect activity

1291 [30] [31] [32]

921

using Landsat MSS,” Photogram. Eng. Remote Sens., vol. 49, 1983, pp. 1303-1314. J. R. Jensen, Introductory Digital Image Processing. Englewood Cliffs, NJ: Prentice-Hall, 1986. J. E. Vogelmann, “Detection of forest change in the Green Mountains of Vermont using multispectral scanner data,” Int. J . Remote Sens., vol. 9, pp. 1187-1200, 1988. S. Khorram, J. A. Brockhaus, R. I. Bruck, and M. V. Campbell, “Modeling and multitemporal evaluation of forest decline with Landsat TM digital data,” IEEE Trans. Geosci. Remote Sens., vol. 28, pp. 746-748, 1990. A. N. Rencz, “Multitemporal analysis of Landsat imagery for monitoring forest cutovers in Nova Scotia,” Can. J . Remote Sens., vol. 1 1 , pp. 188-194, 1985. T. Fung, “An assessment of TM imagery for land cover change detection,” IEEE Trans. Geosci. Remote Sens., vol. 28, pp. 681-684, 1990. W . D. Hudson and C. W. Ramm, “Correct formulation of the Kappa coefficient of agreement,” Photogram. Eng. Remote Sens., vol. 53, pp. 421-422, 1987. J. T. Hahn, R. E. McRoberts, and W. Befort, “Annual forest inventory system (AFIS): Integrating data base techniques, satellite imagery, annual designed sampling and modeling,” in Proc. Conf. lntegrat. Forest Resource Informat. Space, Time, Canberra, Australia, 1992, pp. 314-324.

Pol R. Coppin was born in Kortrijk, Belgium. He received the diploma in agricultural engineering, from the Forestry Department, University of Gent, Belgium in 1977 and the Ph.D. degree in forestry, with an emphasis on remote sensing, from the University of Minnesota St. Paul, in 1991. From 1977 to 1988, he was active as a natural resources assessment specialist in Central and Latin America, Southeast Asia, Europe, and Africa, specifically in integrated survey design, sampling, mapping, aerial photography, digital remote sensing, GIs, applied research, and technology transfer. He continues to participate in international endeavors via short-term consultancies. He is presently an Assistant Professor of biometrics and remote sensing at Purdue University, West Lafayette, IN, where his responsibilities include teaching courses in natural resources assessment, nondigital remote sensing, digital remote sensing, and spatial data processing. His current research interests focus on the quantitative and qualitative monitoring of forest canopy dynamics and land cover changes in the temperate and tropical regions using passive and active remote sensors, the design of spatial information systems to enhance the decision-making processes in natural resources management, hyperspectral remote sensing, and wetland monitoring.

Marvin E. Bauer received the B.S.A. degree in agriculture 1965 and M.S. degree in agronomy in 1967 both from Purdue University West Lafayette, IN, and the Ph.D. degree in agronomy 1970 from the University of Illinois, Urbana. From 1970 to 1983 he was a research agronomist and program leader of crop inventory research at the Laboratory for Applications of Remote Sensing, Purdue University, where he had key roles in the design and implementation of several large scale agricultural remote sensing experiments. Since 1983 he has been a Professor of remote sensing at the University of Minnesota and was director of the Remote Sensing Laboratory from 1983 to 1991. His current research interests include measurements and modeling of the spectral reflectance characteristics of forest stands and canopies in relation to their biophysical and ecological characteristics and the development and application of satellite remote sensing to resource inventory and management. Dr. Bauer is a member of the American Society of Agronomy, the IEEE Geoscience and Remote Sensing Society, the American Society of Remote Sensing and Photogrammetry, and the Electromagnetics Academy-hstitute for Electromagnetics Modelling and Applications. Since 1980 he has served as Editor-in-Chief of the journal Remote Sensing of Environment.