International Journal of Environmental Protection
April. 2013, Vol. 3 Iss. 4, PP. 10-16
Satellite Remote Sensing for Spatio-Temporal Estimation of Leaf Area Index in Heterogeneous Forests Yaseen T. Mustafa Department of Environmental Science, Faculty of Science, University of Zakho, Inter. Road, P. O. Box 12 Duhok, Kurdistan region, Iraq
[email protected] Abstract-Biophysical parameter values such as LAI have proved useful in a number of environmental applications. An approach is presented for producing the spatio-temporal estimation of leaf area index (LAI) of a heterogeneous forest using Moderate Resolution Imaging Spectroradiometer (MODIS) satellite images. This is performed by decomposing MODIS LAI for a heterogeneous forest using the Linear Mixture Model (LMM) and the information about the class fraction from an aerial image. Results showed that the decomposed MODIS LAI values were estimated well with maximum and minimum RMSE of 0.37, and 0.17, respectively. We concluded that our approach can be used to decompose MODIS LAI successfully for any heterogeneous forest. Keywords- Leaf Area Index (LAI); Mixed Pixels; Moderate Resolution Imaging Spectroradiometer (MODIS) Satellite Imagery
I.
INTRODUCTION
Global changes are driving alterations in climate, water resources, and ecological systems. Monitoring these changes, and thereby especially the spatial and temporal variability of vegetation properties, is one of the foremost fields of research within remote sensing. The knowledge of these variables is crucial to understanding terrestrial biosphere processes and can be used for the parameterization of various physical models, quantifying the exchange of energy and matter between the land surface and the atmosphere. Due to the role of green leaves in controlling biological and physical processes in plant canopies, the Leaf Area Index (LAI) is a key element of vegetation structure and an important input parameter to hydrological models. It is defined as half the total leaf surface area per unit ground surface area projected on a local horizontal datum [1]. The LAI can be estimated in the field either by direct (e.g., destructive sampling and litterfall) or by indirect (e.g. measurements of light transmission through canopies) procedures. Although their extensive usage, these methods demand considerable amounts of labour and time especially for a large spatial and temporal variability [2]. Therefore, satellite remote sensing systems have been proposed as a good solution for measurement of forest parameters as providing extensive spatial information for different forest phenomena [3]. For instance, the Moderate Resolution Imaging Spectroradiometer (MODIS) on-board Earth Observing System (EOS) Terra/Aqua platforms, provides LAI as a standard product (MOD15) at a 1 km support, every eight days [3]. The low spatial support of the MODIS LAI product, however, limits the utility at local scales especially in the heterogeneous forest that are mostly planted, as in this study. Mustafa et al. [4] have estimated MODIS LAI at 250 m spatial support using MODIS normalized difference vegetation index (NDVI) product. Even with this finer support, the high landscape heterogeneity can be completely dissolved in 250 m pixels. Consequently, it could be concluded that such areas are not well represented in readily available global data and require more detailed surface representation by high spatial resolution satellite images such as IKONOS. The high spatial resolution images, however, are often not available during the year, due to atmospheric characteristics such as the presence of clouds and aerosols, and those images are expensive. Thus, for effective monitoring of terrestrial ecosystems for a long period of time it is more convenient to use lower cost high temporal resolution data. The aim of this study is to estimate LAI in heterogeneous forests that show variation in space and time using MODIS data. This is performed by decomposing MODIS LAI, thus produce separate data for individual species. The study is applied to a heterogeneous forest, called the Speulderbos, in The Netherlands. II.
MATERIALS AND METHODS
A. Study Area The Speulderbos forest is located between 51o 96′ 06″ to 52o 380 00″ N and 05o 61′ 69″ to 06o 08′ 90″ E (Fig. 1), it was planted in 1962. The tree height in 2006 was approximately 32 m with a density of 785 trees ha-1. The forest cover consists of five tree species, i.e., Beech (Fagus sylvatica), Douglas fir (Pseudotsuga menziesii), Hemlock (Tsuga spp), Japanese larch (Larix kaempferi) and Scotch pine (Pinus sylvestris) [5]. A 46 m high tower with a climate station is situated at 52o 15′ 08″ N, 05o 41′ 25″ E, near the village of Garderen. An area of 4 km2 has been considered in this study, Fig. 1.
- 10 -
International Jouurnal of Environmental Prottection
April.. 2013, Vol. 3 Iss. 4, PP. 100-16
Fig. 1 Maap (a) region of T The Netherlands, (b) high resolutio on aerial image (c) classifiication of forest trree species using manual segmenttation, and (d) sam mpling design of the 1×1 km
B. LAI Ground Data 1) Spatial SSampling: Prior to the ffield work, a natural n colour aerial image with a 0.25 m spatial suppo ort acquired du during the sum mmer of 2011 was w o determine fforest cover sp pecies that veerified in the field. Subsequently, the aeerial useed. A visual iinterpretation allowed us to imaage was mannually segmennted into five segments, onne for each trree species an nd used as a rreference polygon (Fig. 1((c)). Thereafter, the sample points were distrib buted over eaach segment following a stratified s randdom sampling g, Fig. 1(d). The mber of sampple points assiigned to each tree segmentt was based on the proportion to the areea covered by that segmentt. In num totaal 155 samplee points were distributed over the studyy area (Fig. 1((d)). The locattion of each ssample point was w mapped in a fielld campaign uusing two hanndheld GPS devices (Garm min eTrx and MobileMapper M rTM 6) with ±33 m accuracy y. These locatiions weere labelled byy marking thee trees to allo ow taking connsecutive LAII measuremen nts (five timess within every y 16 days) at the sam me locations, ffrom the end of o April to thee end of June 22011. 2) Methodss and Measureements: The Digital Hemisphericaal Photograph hs (DHPs) tecchnique is ussed for LAI measurements m s. We used a Canon EOS 5D cam mera (21 meggapixel) with a 15 mm F2.8 8 EX fisheye llens and a 180 0o field-of-vieew (15 mm, f/ f/2.8). The cam mera with fishheye len ns was held at 1.3 m above the t ground an nd mounted onn the tripod an nd pointed at 90 9 o in the verttical direction n (Fig. 2). At each e sam mple point thrree photos witth a distance of o 5 m were ttaken and processed with th he Gap Light Analyzer (GL LA) software [6]. Alll measuremennts were madee under diffusse light condittions to avoid d introducing errors due to the presence of sunlit foliaage. The LAI field daata (LAIFD) arre obtained fo or all tree speccies of area nu umber 2 as sho own in Fig. 1 (c), and from m April 23rd, 2011 th J 26 , 20111, and used as a validation data. to June
- 11 -
April.. 2013, Vol. 3 Iss. 4, PP. 100-16
International Jouurnal of Environmental Prottection
Fig. 2 Hemispherical camera aand tripod set up in the discontinu uous forest site
C. Satellite Dataa For the sake of simplicity, we assume that t pixels havve a square fo ootprint, that the t ground saampling distan nce (GSD) equuals w ignore the effect of the ppoint spread function f (PSF)). LAI estimattion values caan be providedd by thee size of the foootprint, and we a sp paceborne prooduct, in particular MODIS S every 8-dayss at 1 km spatiial support [7]. The MODISS data sets for this study weere: MOD033: Geolocatioon Fields Daiily Level 1A (L1A) swath h 1 km (origiinally 926.6 m) data, con ntaining geoddetic coo ordinates, groound elevationn, solar and satellite zenithh, and azimuth h angles for each e MODIS 1 km samplee. These data are pro ovided as a ‘coompanion’ daata set to the Level L 1B (L1B B) calibrated raadiance [8]. MOD113Q: The VI product is gridded g maps depicting spatial and tem mporal variatioons in vegetaation activity for mo onitoring of thhe Earth’s terrrestrial vegetaation. The bannds of interestt in this study are day of thhe year (DOY), the NDVI, and pix xel reliability. NDVI data arre available on a 16 day baasis at a 250 m spatial support, whereas tthe original prroduct is of 2331.7 m support s that iss used in this study. The pix xel reliability band in the NDVI N product describes if ppixels were geenerated usingg the hisstorical filling criteria [9]. MOD155A: MODIS LAI/ L Fraction n of Absorbedd Photosyntheetically Activee radiation (FPPAR) product. It is compoosed eveery 8 days at a 1 km spatiaal support on a Sinusoidal ((SIN) 10-degrree grid [7]. In I this study tthe original prroduct supporrt of 926 6.6 m is used.. The globe is tiled into 36 tiles along thee east–west ax xis, and 18 tilees along the nnorth-south ax xis, each coverring 2 120 00×1200 km [8]. The compposite 8 day product p is prodduced by using maximum FPAR. F In [4], we deerived LAI froom the MODIS NDVI produ duct at a nomin nal 231.7 m sp patial supportt and composiited over 16 daays. This is used in thhis study to prroduce LAI att finer supportt and we referrred to it as a modified m MOD DIS LAI (LAIM) to distinguuish f the MOD DIS LAI produuct provided by b NASA. it from D. MODIS LAI Decompositioon 1) Determinnation of Subppixel Area Pro oportion: The LAIM is derived from m the compositted MODIS N NDVI productt of 231.7 m every e 16 days.. A filter baseed on the MOD DIS v geom metry resulted in a compositte VI image coontaining pixeels from differrent VI algorithm inccluding qualityy, cloud, and viewing ys of the yearr. The DOY used in the composite c prooduct is linkeed to the L1B B footprint off that day, ideentified using the day infformation from m the composiite DOY band d. The Level 4 MODIS globbal LAI/FPAR R product is aassumed to have h a fixed grid g all the tim me with the spatial s supporrt of 926 6.6 m [7, 8] aand labelled as a a reference footprint. Foour reference footprints are selected thatt fit to the sizze of the studyy of inteerest. These reeference footpprints are num mbered as 1, 2,, 3, 4 (Fig. 1 (b), ( and (c)) an nd gridded intto 16 cells of spatial supporrt of 231.7 m to matcch with the MODIS NDVI product. In orrder to identiffy the correct geolocation oof each pixel in i the MODIS S VI pro oducts, we ussed MODIS MOD03 M geolo ocation data oof the day av vailable from the composititing VI imag ges. The MOD D03 imaages with a sppatial support of 926.6 m arre also griddedd into 16 cellss of 231.7 m. MOD03 M griddded cells of the identified DOY D aree intersected w with the griddeed cells of refe ference footpriint and a shorttest distance between b centeers of a grid ceell of the MOD D03 foo otprint and thhe grid cell of o the reference footprint iis selected (F Fig. 3 (a)). The selected ggrid cell of MOD03 M footpprint con nsidered to bbe the correspponded groun nd cell. Subseequent, the area a proportio on is determinned by interssecting refereence pollygons of eachh tree species with the grid d cell of MOD D03 footprint of o 231.7 m su upport. Fig. 3 (b) illustrates the procedure of dettermining the area proportioon of each speecies within thhe MODIS griid cell that corrresponds to th the ground celll. This proceddure waas applied to alll composite MODIS M NDVI images (5 im mages) over th he study time of two monthhs.
- 12 -
International Jouurnal of Environmental Prottection
April.. 2013, Vol. 3 Iss. 4, PP. 100-16
Fig. 3 (a) An exxample of determiining a proper griid cell of MOD033 by selecting a shortest distance between b the centrre of a grid cell (× ×) of the MOD03 footprint and the t centre of a griid cell ( )٭of the reference r footprin nt (dashed line) (b) Floowchart process of o determining arrea proportion of each species with hin the MODIS ggrid cell
2) Decomposition: We used the Linear Mixtuure model (LM MM) to decom mpose the LAIIM value. The LMM assum mes that the spectrum measuured by a sensor is a linear combinnation of the spectra s of all ccomponents within w the pixel [10]. Accorrding to its deefinition, the LAI L oveer a specific aarea can be innterpreted as a linear compoosition of LAIs in subareass. This is achiieved by considering the LA AIM pix xel as a linearr combinationn of LAI valu ues of all speccies present in n the modified MODIS pixxel, and we assumed a one LAI L vallue of a certainn species in thhe entire study y area by ignooring the spatial variation off the species. T The LMM is thus t expressedd as: ∑
,
1
where i (i = 1, …, 5) referrs to the imagee number, is the LAIM for f pixel k at image i, A is a matrix of th he area proporttion vallues of speciess j (j = 1, …, 5) in pixel k at a image i, is the LAI of tree species j at image i, annd is the reesidual for pixxel k at image i i. , vvalues are varrying over imaages i based oon the size of the polygon of o each speciees j occupies in the footprinnt of grid cell of MOD03. Therefoore, the explan natory variablles in Expresssion (1) are th he xi (LAIM) oof each pixel and the Ai in the grid cells of MO OD03. Hencee, solving (1) using least sqquares leads to retrieving LAI values oof each speciees in the MOD DIS imaage that is refferred as the decomposed d MODIS M LAI ((LAISAT). The resulting resiidual values bby LMM for each e species were w useed to assess thhe fit of the model. m We app plied this proccedure to the set of MODIS S images thatt were collecteed between April A r 23rd , 2011 to Junne 26th, 2011,, by assuming g that there is no major change in the forrest compositiion as labeled d by the refereence pollygons of eachh species. i
III. RESULTS Fig. 4 showss the resulted LAISAT valuees by decompposing the LA AIM values. We W noticed uneexpected deco omposed MOD DIS LA AI values at Juune 10th, 2011. This was due to MODIS oorbit observattions being farr from nadir aat all 16 days for f the compoosite MO ODIS NDVI. This made thee footprint ressolution of a ggrid cell more than 1.2 km. Consequentlyy, the area pro oportion resultting fro om intersectingg the referencce polygons with w a grid celll footprint inffluences the calculation of tthe LAI valuees for all speccies. The unexpectedd decomposedd MODIS LA AI value of alll species is replaced with the precedennt values (Fig g. 5). The spaatiotem mporal distribuution for areaa number 2 (Fiig. 1 (c)) of thhe LAISAT at the t spatial sup pport 231.7 m of mixed species is shownn in Fig g. 6. This is peerformed afterr considering the proportionn of each speccies in the com mposition of eeach pixel. Th he accuracy off the deccomposed LA AI was tested using u the Root Mean Squarre Error (RMS SESAT) and thee Relative Errror (RESAT) rate with respecct to thee LAI field daata. We foundd that the outccome of the ddecomposition n approach waas satisfactoryy with RMSESAT of 0.31, 0.26, S 0.3 37, 0.17, and 00.23 for Beecch, Douglas fiir, Hemlock, JJapanese larch h, and Scotch pine, respecttively (Table I). I Moreover, the RE ESAT of all of B Beech, Douglas fir, Hemlocck, Japanese llarch, and Sco otch pine weree 7.31%, 6.433%, 8.68%, 3.47%, and 4.477%, resspectively (Tabble I).
- 13 -
International Jouurnal of Environmental Prottection
April.. 2013, Vol. 3 Iss. 4, PP. 100-16
Fig. 4 D Decomposed MO ODIS LAI into fiv ve LAI values of five species, Beeech, Douglas fir, Hemlock, H Japanesse larch, and Scotch pine
Fig. 5 Decom mposed MODIS LAI L into five LA AI values of five sspecies after replaacing the unexpec cted LAI values w with the preceden nt LAI values. (a) Beech, (b) Douglas fir, (c) H Hemlock, (d) Jap panese larch, and (e) Scotch pine
Fig g. 6 The spatio-tem mporal distributioon of the mixturee LAI values conssidering the area proportion p of eac ch species from ennd of April 2011 till end of June 2011 2
- 14 -
International Journal of Environmental Protection
April. 2013, Vol. 3 Iss. 4, PP. 10-16
TABLE I THE MEAN VALUES (), THE STANDARD DEVIATION (), ROOT MEAN SQUARE ERRORS (RMSESAT) AND RELATIVE ERRORS (RESAT) FOR THE DECOMPOSED LAI (LAISAT), AND THE LAI FIELD DATA (LAIFD) OF FIVE TREE SPECIES DURING STUDY TIME
Beech Douglas fir Hemlock Japanese larch Scotch pine Composed LAI
Source LAISAT LAIFD LAISAT LAIFD LAISAT LAIFD LAISAT LAIFD LAISAT LAIFD LAISAT LAIFD
4.24 4.09 4.24 4.00 4.27 4.04 4.13 4.05 4.31 4.16 4.17 5.60
0.22 0.43 0.20 0.20 0.36 0.54 0.11 0.26 0.18 0.35 0.20 0.43
RMSESAT 0.31
RESAT 7.31%
0.26
6.43%
0.37
8.68%
0.17
3.47%
0.23
4.47%
1.45
25.25%
IV. DISCUSSION The LAI value for a heterogeneous forest that shows variation in space and time is estimated using MODIS data. It is achieved by using LMM and information of the area proportion of each species within pixels of MODIS images. Our results showed that the decomposed MODIS LAI is estimated with significant values. The deviation between decomposed MODIS LAI and LAI field data are lower than the deviation between the composed MODIS LAI and LAI field data. The strength of this approach is that the spatial LAI values can be estimated for a heterogeneous forest by classifying satellite images into the forest cover species. It requires dealing with the original raw data of satellite images and the use of the Linear Mixture Model. This study showed the applicability of the method with several tree species by treating each species individually and combines them taking in the consideration the area proportion of each species. Hence, a reliable and accurate LAI values can be estimated for mixture pixels. We realize that the decomposition of MODIS LAI and identifying area proportions of each tree species directly using the nominal 250 m support of the VI product is not a suitable procedure. This is because the VI product is not a realistic representation of the actual ground area (i.e., instantaneous field of view (IFOV)) measured by the MODIS sensor. The composite VI images are derived by compositing image data from multiple images acquired from different orbit paths over a 16-day period, which has a different footprint every time. This was a reason of determining area proportion using MOD03 footprints. Two assumptions were proposed to implement the decomposition approach. First, no major change in the forest composition in terms of area proportion in the reference polygons of each species occurred during the study period. Second, one LAI value assumed for a specific species by ignoring a spatial variation of the species and considered as an average value. The latter assumption rarely holds in practice, but it allows a simple unmixing. This is, however, limiting the approach and a more dedicated model is required to do this species unmixing more reliably. These assumptions as sources of uncertainty might influence the accuracy of the LAISAT. Some issues require further work. Extension of the time period and increased of the spatial scale may show a better explanation of seasonal changes in the LAI. Also, ground measurements should be collected for the study area at the same spatio-temporal support of the satellite images might improve validation of our approach. Spatial data of the forest are becoming increasingly important for forest vegetation management, decision making and for climate change issues. The reliable and accurate estimations of the biophysical variables as LAI are useful and have served as inputs to functional models of ecosystem biogeochemistry. It provides a better understanding of the forest, description of the spatial pattern of forest structure, and which ultimately can serve as an informative factor in climate change. V.
CONCLUSIONS
In this study the LAI estimation for individual species with MODIS pixels is achieved by delineating the area proportion of each species followed by LMM. We conclude that the decomposed MODIS LAI of individual species is estimated successfully, such that the deviation between the decomposed LAI and the LAI filed data is 74.48% lower than the deviation between composed MODIS LAI and the LAI field data. Lastly, given the presented method and results in this study, our approach can be applied successfully to estimate LAI values for any heterogeneous forest. ACKNOWLEDGMENT
The author would like to thank Dr. V. A. Tolpekin for his effort and support especially in programming. REFERENCES
[1] D. J. Watson, “Comparative physiological studies on the growth of field crops. I. Variation in net assimilation rate and leaf area between species and varieties, and within and between Years,” Annals of Botany, vol. 11, pp. 41-76, 1947. [2] M. Weiss, et al., “Review of methods for in situ leaf area index (LAI) determination Part II. Estimation of LAI, errors and sampling,”
- 15 -
International Jouurnal of Environmental Prottection
April.. 2013, Vol. 3 Iss. 4, PP. 100-16
Agric. For. M Meteorol., vol. 121, 1 pp. 37-53, Jan 2004. [3] F. Gao, et al., “An Algorithm m to Produce Temporally T andd Spatially Conttinuous MODIS S-LAI Time Serries,” IEEE Geosci. Remote Sens. S Lett., vol. 5, ppp. 60-64, 20088. [4] Y. T. Mustaffa, et al., “Bayesian network modeling m for im mproving forest growth estimattes,” IEEE Tranns. Geosci. Rem mote Sens., vol.. 49, pp. 639-649, 2011. [5] W. Timmerm mans, et al., “EA AGLE 2006-EA AGLE Netherlaands Multi-purp pose, Multi-Ang gle and Multi-ssensor In-situ, Airborne A and Sppace Borne Campaigns over Graassland and Forrest, Final repoort,” Proceeding gs of the AGR RISAR and EAG GLE campaign ns, Noordwijk, The Netherlands, January 2008. [6] J. P. Hardy, eet al., “Solar raddiation transmisssion through coonifer canopiess,” Agric. For.M Meteorol., vol. 1126, pp. 257-27 70, Nov 2004. v leaf aarea and fractio on absorbed PAR from year onne of MODIS data,” d Remote Sens. S [7] R. B. Mynenii, et al., “Globaal products of vegetation Environ., voll. 83, pp. 214-2331, Nov 2002. orage, griddingg, and compositting methodology: Level 2 griid,” IEEE Tran ns. Geosci. Rem mote [8] R. E. Wolfe, et al., “MODIIS land data sto Sens., vol. 366, pp. 1324-1338, 1998. Solano, K. and a Jacobson, A. and Huuete, A. (201 10). MODIS Vegetation Inndex User's Guide. Availaable: [9] R. a. D. S http://vip.arizzona.edu/docum ments/MODIS/M MODIS_VI_UssersGuide_01_2 2012.pdf.
[10 0] M. Brown, ett al., “Linear sppectral mixture models m and suppport vector maachines for remo ote sensing,” IE EEE Trans. Geo osci. Remote Seens., vol. 38, pp. 22346-2360, 20000. Yaseen T. T Mustafa recceived the B.Scc. and M.Sc. degrees d in math hematics from the University y of Mosul andd the Universityy of Duhok Iraq q in 2000 and 22005, respectively, and the Ph.D. degree in ap applied statisticss in remote sennsing from Facuulty of Geo-Infformation Sciennce and Earth Observation (IT TC) of the Uniiversity of Tweente, Enschede,, the Netherlannds in 2012. Currently, he is a lecturer in the t Faculty of Science at thee University off Zakho, Kurdiistan Region-Irraq. His researcch interests are the area of rem mote sensing, in ncluding matheematical and staatistical tools, such s as Bayesian networks. Applications A eme merge from a ran nge of agricultural, urban, and environmental fields. Dr. Musstafa E, the Americann Society for Ph hotogrammetry and Remote Seensing (ASPRS), and International is a membber of the IEEE Society foor Environmenttal Information Sciences (ISEIIS).
- 16 -