Mapping spatial variation in acorn production from ... - Springer Link

2 downloads 0 Views 558KB Size Report
Mapping spatial variation in acorn production from airborne hyperspectral imagery. YAO Zhong *, Kenshi SAKAI. Institute of Symbiotic Science and Technology, ...
For. Stud. China, 2010, 12(2): 49–54 DOI 10.1007/s11632-010-0010-9

RESEARCH ARTICLE

Mapping spatial variation in acorn production from airborne hyperspectral imagery YAO Zhong *, Kenshi SAKAI Institute of Symbiotic Science and Technology, Department of Ecoregion Science, Faculty of Agriculture, Tokyo University of Agriculture and Technology, Tokyo 183-8509, Japan

© Beijing Forestry University and Springer-Verlag Berlin Heidelberg 2010 Abstract Masting is a well-marked variation in yields of oak forests. In Japan, this phenomenon is also related to wildlife management and oak regeneration practices. This study demonstrates the capability of integrating remote sensing techniques into mapping spatial variation of acorn production. The hyperspectral images in 72 wavelengths (407–898 nm) were acquired over the study area ten times over a period of three years (2003–2005) during the early growing season of Quercus serrata using the Airborne Imaging Spectrometer Application (AISA) Eagle System. With the canopy spectral reflectance values of 22 sample trees extracted from the images, yield estimation models were developed via multiple linear regression (MLR) analyses. Using the object-oriented classification approach in eCognition, canopies representative of individual oak trees (Q. serrata) were identified from the corresponding hyperspectral imagery and combined with the fitted estimation models developed, acorn yield over the entire forest were estimated and visualized into maps. Three estimation models, obtained for June 27 in 2003, July 13 in 2004 and June 21 in 2005, showed good performance in acorn yield estimation both for the training and validation datasets, all with R2 > 0.4, p < 0.05 and RRMSE < 1 (the relative root mean square of error). The present study shows the potential of airborne hyperspectral imagery not only in estimating acorn yields during early growing seasons, but also in identifying Q. serrata from other image objects, based on which of the spatial distribution patterns of acorn production over large areas could be mapped. The yield map can provide within-stand abundance and valuable information for the size and spatial synchrony of acorn production. Key words yield map, estimation model, classification map, acorn, spatial synchrony, hyperspectral imagery, masting

1

Introduction

In many trees, crop production alternates between heavy and light harvests and flowering and fruit production are often synchronized over extensive spatial areas (Kelly, 1994; Koenig and Knops, 1998; Koenig et al., 1999). This phenomenon is generally referred to as alternate bearing, also known as masting or mast fruiting in acorn production. In Japan, fluctuation in acorn production affects the migration of wild boars, which may have destructive effects on farms and even local residential areas. Furthermore, acorns are a valuable seed source for forest restoration (Takahashi et al., 2007; Yasaka et al., 2008). In years of low acorn production, most seeds are consumed by wildlife, which may be an important factor in oak regeneration failure in many areas (Galford et al., 1991). Therefore, to obtain information on acorn crop size and spatial distribution patterns and to understand the masting mechanism in acorn production is of fundamental importance both to oak regeneration practices and wildlife management. Yield maps serve as location-year records of spatial variation in crop production and more specifically, *

Author for correspondence. E-mail: [email protected]

may be the trajectory of food sources for much of the wildlife in forest ecosystems, which may be predictive of foraging-induced wildlife assemblage. In agriculture with its commercialization of combine-mounted grain yield monitors, the creation of accurate yield maps has been an essential component of the successful implementation of precision farming (Uno et al., 2005). The advent of remote sensing offers a potentially attractive alternative to combine-mounted yield monitors, since a number of studies have demonstrated the ability of remote sensing in estimating crop yields (Yang et al., 2004; Dente et al., 2008). In forests, the rapid and cost-effective application of remote sensing to map forest stands and quantitative parameters such as leaf area index and stem volume, is a consistent motivation for its utilization in forest management planning to replace more labor-intensive, time-consuming and mostly post-harvest field inventories (Dymond et al., 2002; Clark et al., 2004; Souza et al., 2005; Chambers et al., 2007). However, so far studies on the application of remote sensing to yield monitoring have mainly been conducted on agricultural crops. Few studies have been carried out on forest trees, probably due to the complexity of forest

50

Forestry Studies in China, Vol.12, No.2, 2010

ecosystems. Our research was conducted as part of a project to determine the capability of hyperspectral imagery in investigating spatial variation in acorn production. This presentation reports on our attempts to estimate acorn yields over large areas based on yield estimating models previously developed from airborne hyperspectral imagery. As well, we analyzed the scale at which masting occurs based on spatial distribution patterns of estimated acorn production from yield maps.

2 2.1

Materials and methods Study area

This research was conducted in an experimental acorn forest located at the Field Science Education Research Center, affiliated with Tokyo University of Agriculture and Technology, Hachioji City, Japan (35°38′N, 139°23′E; Fig. 1A), with an area of approximately 0.65 km2. This area has a temperate climate, with an annual mean minimum temperature of 9.4°C and annual precipitation of 1303.3 mm in 2003; corresponding data for 2004 were 10.1°C and 1570.4 mm (in October given the effect of Typhoons 22 and 23) and 8.6°C and 998.3 mm for 2005. Quercus serrata, also known as konara in Japan, is a deciduous broadleaf tree species belonging to the family Fagaceae. It is a common dominant in secondary forests in Japan. In our study area, Q. serrata is the major dominant canopy species, with other arboreal species such as Japanese cedar (Cryptomeria japonica), bamboo (Phyllostachys bambusoides) and pines (Pinus densiflora and P. thunbergii) scattered at lower densities. The seeds of Q. serrata are very important for the behavior of local wild boars and black bears. Twenty-two trees were used as sample trees to develop the estimation models in our study (Fig. 1A).

2.2

Hyperspectral data

An Airborne Imaging Spectrometer for Applications (AISA), i.e., the Eagle system (Pasco Co. Ltd., Tokyo, Japan) was used to obtain hyperspectral images over the study area in three consecutive years (2003, 2004 and 2005), with ten flights during the early growing season (April 10, May 22, June 5 and June 27 in 2003, May 25, June 18 and July 13 in 2004, May 26, June 21 and July 21 in 2005). Images were collected with 1.5 m × 1.5 m spatial resolution, 72 channels (from 407 to 898 nm) and 6.3 nm spectral resolution, taken at an altitude of approximately 1000 m above the ground during cloud-free periods in the daytime. The hyperspectral data were processed to at-sensor radiance, using calibration coefficients determined in the laboratory by SPECIM (Spectral Imaging Ltd., Oulu, Finland). An onboard Fiber Optic Downwelling Irradiance System (FODIS) was used to transform at-sensor radiance to surface reflectance. The images were atmospherically calibrated using the modified flat field method. Spectra of a sample field were obtained using a spectrophotometer during the flight of the AISA Eagle on each date, which were then used for atmospheric calibration. The distortion/geometry of images was corrected for aircraft movements (yaw, pitch and roll) using on-board Global Positioning System/Inertial Navigation System (GPS/INS) data. The images were then rectified to Universal Transverse Mercator (UTM) geographic coordinates. Further rectification was performed using a reference map while keeping the estimated root mean square error (RMSE) at less than 0.5 pixels. 2.3

Estimation model based on PLS B-Matrix

To obtain hyperspectral data of 22 sample-trees for the estimation models, corrected images were imported

Fig. 1 Acorn forest at the Field Science Education Research Center, Hachioji City, Japan (polygons in blue indicate 22 sample trees) (A), and example of canopy spectral reflectance data for the 22 sample trees from the hyperspectral imagery (B)

YAO Zhong and Kenshi SAKAI: Mapping spatial variation in acorn production from airborne hyperspectral imagery

was employed to test for consistency in the models developed. For details, see our previous paper (Yao et al., 2008). Data analysis was performed in a MATLAB environment. The code used for PLS was from the n-way Toolbox 2.11 for use with MATLAB (Andersson and Bro, 2000). Other statistical analyses were carried out based on codes from the Statistics Toolbox in MATLAB.

into the ERDAS IMAGINE software (ERDAS IMAGINE 8.6), where individual tree canopies were manually identified by drawing a polygon over the image objects that corresponded to their respective trees on the ground. The locations for individual sample trees were determined by the Handy Global Positioning System and polygon-shaped canopies were identified from field-pictures. Using embedded hyperspectral data extraction procedures in the software, the mean reflectance values of each canopy at each wavelength for each season were abstracted and used as original hyperspectral data for model development (Fig. 1B). Acorns from individual sample-trees were collected each year from 2003 to 2005 using conventional seed-traps. Two to five seed-traps were set underneath each of the 22 sample trees according to the size of the canopy of the sample trees in April and were withdrawn in December. Individual yield data (number of acorns) were then obtained as sum of the number of acorns divided by the number of traps set. The original yield data were log-transformed as original yield information for model development. Considering that there may be multicollinearity and redundancy in hyperspectral data (Broge and Leblanc, 2001) the B-matrix technique, based on partial least squares (PLS) analysis, was applied to determine the important wavelengths as the estimating factors. The B-matrix can be calculated from the PLS loadings and weights as follows:

B = W ( P TW ) −1 Q T

2.4

Acorn yield map

Before creating the yield map, it was necessary to identify the Q. serrata from other objects in the images, i.e. the species-level classification. Considering the canopy-based estimation model as well as the heterogeneous forest-covers, we used the eCognition Elements 4.0 (Definiens Inc., Germany) to conduct the classification. However, this software uses a fuzzy rule base, not to classify single pixels, but rather to classify image object primitives, extracted in a previous image segmentation step (Definiens, 2003). Compared to classic hard classifiers (e.g., maximum-likelihood, minimum-distance or parallelepiped), this kind of soft classifier enables detailed performance analyses and provides insight into the class mixture for each image object by integrating its contextual spatial information (Definiens Inc., Germany) and thus provides more detailed and accurate mapping products. In order to create the classification map, the corrected images were imported into eCognition. A network of relatively homogeneous image regions (termed objects) was then generated with a multiresolution segmentation procedure. Based on the synthesized information contained within respective image-objects, the standard nearest neighbor classifier was used to classify the image objects. The process allows a class hierarchy to declare sample objects for each respective class by their manual selection. An example of the sample selection information for image classification is shown in Table 1. As can be seen, except for the two classes of konara and other trees,

(1)

where W denotes predictor-variable weights, P the predictor-variable loadings and Q the response-variable loadings. Using these factors, estimation models based on multiple linear regression (MLR) algorithms were developed, which can be simply expressed as: Y = f(X)

51

(2)

where Y is the log-transformed acorn yield, X estimating factors, some of the most important wavelengths identified, based on the PLS B-matrix and f the multiple linear regression (MLR) function. With respect to the overfitting problem, a cross validation procedure

Table 1 Class separation distance matrix resulting from within the optimal feature space for image acquired on July 13, 2004 Class

65 out of 144 (spectral mean values and standard deviation) Konara

Other trees

Konara

0

0.891

Other trees

/

0

Cedar

/

/

Building

/

Soil

/

Grassland

/

Cedar

Building

Soil

Grassland

2.847

6.242

2.715

1.264

5.944

7.724

4.296

1.932

0

7.687

4.157

4.226

/

/

0

3.065

6.333

/

/

/

0

3.152

/

/

/

/

0

Note: The distance matrix was computed from the normalized Euclidean distance (Definiens, 2003), and the higher the distance, the more separability is between the two classes.

52

Forestry Studies in China, Vol.12, No.2, 2010

other class-pairs show desirable separation distances, indicating the separability between each other. Considering the overwhelming dominance of the species konara in the studied forest, the knowledge base could be used to classify all image objects into their respective class. After the classification process, the mean reflectance values at each wavelength for each output cell (segmented objects) of the class konara, in text file (.asc) from eCognition, were entered into the respective desirable estimating model previously developed for each of the three years (2003–2005) in the MATLAB environment to predict the acorn yield over the entire forest. After the respective classification map in shape file (.shp) from eCognition was imported into ArcGIS 9.1 (ESRI Inc., 2005) and integrated with the estimated yield data, acorn yield maps based on classification images were then generated and visually spread out in the ArcGIS environment.

which part relatively small yields and for each of the three years which year a relatively heavy yield and which year a relatively small yield. Generally, 2003 was a mast year for acorn production in this forest, while 2004 was a non-mast year and 2005 a par year (Fig. 3). In other words, acorn production in this forest showed some degree of synchronization. 3.3 Comparison of yield maps based on pixels and image-objects

We identified the best fitting model for each of the 10 sampling periods across the three years. Of these, the models for June 27 in 2003, July 13 in 2004 and June 21 in 2005 performed in a consistent satisfactory way both for the training and validation datasets (all R2 > 0.4, p < 0.05 and RRMSE (the relative root mean square of error) < 1), as shown in Table 2. Each of these three models showed satisfactory performance, indicating their capability of estimating acorn yield from airborne hyperspectral imagery data.

Pixel-based classification maps and consequent yield maps are very helpful for comparisons, because of the relative homogeneity in each pixel and the range of deterministic information. Therefore, the pixel-based yield maps are also displayed, as shown in Fig. 4. However, this advantage will be obscured when the size of material on the ground is larger than the pixel size or their shape is too irregular to cover the pixel combination. In the study area, the minimum canopy size among the 22 sample trees is 95.01 m2, much larger than the pixel size (ranging from 1 m2 to 4 m2). Moreover, there is the inevitable spectral variation within forest land cover. Therefore, an increase in the signal-to-noise ratio in classification and subsequent yield mapping can be expected when image objects are taken as basic information carriers. Furthermore, as displayed in Figs. 3 and 4, object-based yield maps show much clearer distribution patterns of acorn production (aggregation effects). Besides, in terms of masting phenomena occurring in group behavior, it is more desirable to adopt object-based yield maps where image objects can be regarded as representing individual trees (crown).

3.2

4 Conclusions

3

Results and discussion

3.1

Yield estimating models

Acorn yield map

By applying the estimation models to the classification maps, we created the yield maps based on image objects. Figure 2 illustrates the classification map and the subsequent yield map for each of the three years (2003–2005), corresponding to one of the three fitted estimation models. Objects colored in deeper red indicate high yield production. To some extent, the yield maps visually indicate for each year which part show relatively heavy yields,

We demonstrated the potential and utility of estimating acorn yield from airborne hyperspectral imagery data. Using the B-matrix technique based on PLS analysis to determine the estimating factors, we developed estimation models for 10 seasons from 2003 to 2005 and obtained, respectively, one satisfactory estimation model for each of the three years. Applying these fitted estimation models to the classification map, generated by using the object-oriented classifica-

Table 2 Regression statistics of fitted models for each of three years (2003–2005) Date

Training dataset

Validation dataset

No. of factors

R2

p

RRMSE

R2

p

RRMSE

June 27, 2003

7

0.696

0.001**

0.551

0.599

0.005**

0.673

July 13, 2004

9

0.470

0.020 *

0.728

0.443

0.025*

0.752

June 21, 2005

4

0.719

0.001**

0.530

0.403

0.036*

0.922

2

Note: R and p-values were obtained using a simple linear regression model for the relationship between actual and estimated yields. * indicates significance at p < 0.05 and ** at p < 0.01.

YAO Zhong and Kenshi SAKAI: Mapping spatial variation in acorn production from airborne hyperspectral imagery

53

Fig. 2 Classification maps and yield maps for each of three years (2003–2005). A: original image in RGB format (Red = 660.85 nm; Green = 528.98 nm; Blue = 464.71 nm); B: classification map (only the class Q. serrata); C: yield map.

Fig. 3 Acorn yield map created from object-based classification map

Fig. 4 Acorn yield map created from pixel-based classification map

tion approach in eCognition, we created the yield map. We make the following inferences. 1) The three best fitted models, obtained for June 27 in 2003, July 13 in 2004 and June 21 in 2005, showed

good performance in acorn yield estimation both for the training and validation datasets, with R2 > 0.4, p < 0.05 and RRMSE < 1, indicating the ability of estimating acorn yield from airborne hyperspectral data.

54

2) Using remote sensing data, acorn production in a forest (or on a large scale) can be estimated and visualized into maps, with a background of hyperspectral images, with great savings in time and efforts dedicated to annual yield surveys. 3) Given the yield maps, the variation in both seasonal and spatial acorn production can be visually detected; this shows some degree of synchronized production. 4) Compared to a pixel-based approach, object-based classification maps and the subsequent yield maps show considerably greater success in detecting spatial distribution of acorn production.

Acknowledgements This research was supported by the Japan Society for the Promotion of Science (JSPS) through its grant-in-aid for scientific research projects (No. 14360148). We express our appreciation to Dr. Bhuweneshwar P. Sah and Mr. Suhama of the Pasco Co. LTD., Japan, for the provision of the hyperspectral images.

References Andersson C A, Bro R. 2000. The N-way Toolbox for MATLAB. Chemometr Intell Lab Syst, 52(1): 1–4 Broge N H, Leblanc E. 2001. Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of green leaf area index and canopy chlorophyll density. Remote Sens Environ, 76: 156–172 Chambers J Q, Asner G P, Morton D C, Anderson L O, Saatchi S S, Espírito-Santo F D B, Palace M, Souza C. 2007. Regional ecosystem structure and function: ecological insights from remote sensing of tropical forests. Trends Ecol Evol, 22(8): 414–423 Clark M L, Clark D B, Roberts Dar A. 2004. Small-footprint lidar estimation of sub-canopy elevation and tree height in a tropical rain forest landscape. Remote Sens Environ, 91: 68–89 Definiens. 2003. User Guide 3. Definiens Imaging GmbH, www.definiens-imaging.com Dente L, Satalino G, Mattia F, Rinaldi M. 2008. Assimilation

Forestry Studies in China, Vol.12, No.2, 2010 of leaf area index derived from ASAR and MERIS data into CERES-Wheat model to map wheat yield. Remote Sens Environ, 112: 1395–1407 Dymond C C, Mladenoff D J, Radeloff V C. 2002. Phenological differences in Tasseled Cap indices improve deciduous forest classification. Remote Sens Environ, 80: 460–472 Galford J R, Auchmoody L R, Smith H C, Walters R S. 1991. Insects affecting establishment of northern red oak seedlings in central Pennsylvania. In: McCormick L H, Gottschalk K W (eds). Proceedings of the 8th central hardwood forest conference. Radnor, PA: Northeastern Forest Experiment Station, 271–280 Kelly D. 1994. The evolutionary ecology of mast seeding. Tree, 9: 465–470 Koenig W D, Knops J M H. 1998. Scale of mast seeding and tree-ring growth. Nature, 396: 225–226 Koenig W D, Knops J M H, Carmen W J, Stanback M T. 1999. Spatial dynamics in the absence of dispersal: acorn production by oaks in central coastal California. Ecography, 22: 499–506 Souza C M, Roberts Dar A, Cochranea M A. 2005. Combining spectral and spatial information to map canopy damage from selective logging and forest fires. Remote Sens Environ, 98: 329–343 Takahashi K, Sato K, Washitani I. 2007. Acorn dispersal and predation patterns of four tree species by wood mice in abandoned cut-over land. Forest Ecol Manage, 250(3): 187– 195 Uno Y, Prasher S O, Lacroix R, Goela P K, Karimi Y, Viau A, Patel R M. 2005. Artificial neural networks to predict corn yield from Compact Airborne Spectrographic Imager data. Comput Electron Agric, 47: 149–161 Yang C H, Everitt J H, Bradford J M. 2004. Airborne hyperspectral imagery and yield monitor data for mapping cotton yield variability. Precis Agric, 5: 445–461 Yao Z, Sakai K, Ye X, Akita T, Iwabuchi Y, Hoshino Y. 2008. Airborne hyperspectral imaging for estimating acorn yield based on PLS B-matrix calibration technique. Ecol Inform, 3(3): 237–244 Yasaka M, Takiya M, Watanabe I, Oono Y, Mizui N. 2008. Variation in seed production among years and among individuals in 11 broadleaf tree species in northern Japan. J Forest Res, 13(2): 83–88 (Received December 22, 2009 Accepted February 26, 2010)

Suggest Documents