three cover sampling methods were used, namely, visual estimation, digital photography and 2-D ... has been developed using a digital still-frame video camera ...
IMPROVING GROUND TRUTHING FOR INTEGRATING REMOTELY SENSED DATA AND GIS* Qiming Zhou
Petter Pilesjö
School of Geography University of New South Wales Sydney 2052, AUSTRALIA
Remote Sensing and GIS Lab Department of Physical Geography University of Lund S-221 00 Lund, SWEDEN
ABSTRACT Accurate field estimation of vegetation cover is essential for any improvements in rangeland vegetation modelling and monitoring. This paper reports a recent research project which focuses on improved field techniques for rangeland vegetation investigation. A semi-arid rangeland field station has been selected for intensive testing of a number of field techniques employed by pasture management and remote sensing. The field sampling results from visual estimation, line intercept, quadrat and plant 2-D crown cover model have been compared with digital images which were acquired from a digital camera mounted on a five-metre high tripod. The statistical analysis of the sampling results has shown that visual estimation methods, which are commonly used in rangeland remote sensing, delivered most unreliable and inconsistent results. The study also shows that a high-level of errors has been associated with data aggregation while performing rangeland vegetation ground truthing. A new method has been invented using a remotely controlled digital camera with associated image interpretation methods which allow accurate and consistent estimates of vegetation cover.
1.0 INTRODUCTION Rangeland constitutes 74% of Australian continent and maintaining the land in a sustainable condition has been drawing great attention of Australian public. Extensive livestock production is the major land use on the rangelands with large areas of land required per head of livestock. The rangelands support large proportion of Australia’s cattle and sheep production which are mainly for export sale, thus providing substantial input into the domestic economy (Frazier, et al., 1994). Management of rangelands is ecological in nature, of a low energy input and involving actions that seek to modify, rather than control, the natural forces operating on the land. To maintain the rangelands in a productive condition, a land manager must be able to monitor conditions at a level commensurate with the large areas involved, typically low productivity per unit area, and constraints of labour availability. Remote sensing technology is recognised as an appropriate tool for rangeland management given that it is capable of providing information on the state of vegetation on a regular, continuous and near real-time basis. However, to be useful for land management, remotely-sensed data must be processed and interpreted in the context of geographical and management information. It is now widely accepted that
* in Proceedings of International Workshop on New Developments in Geographic Information Systems, 6-8 March 1996, Milan, pp 30-38.
Geographical Information Systems (GIS), when incorporated with remotely-sensed data, can offer farm managers a useful tool for monitoring environmental conditions and allocating livestock in a spatially efficient manner under rangeland conditions. This is particularly important to assist farmers to manage their land during ‘critical’ periods, eg. during severe droughts which have now become a national-wide environmental and economic concern. One of the obstacles that limit remote sensing application to rangeland farm management is the poor computational models which link spectral characteristics recorded by remote sensors to vital parameters for rangeland management, such as quantitative measurements of vegetation cover and biomass. This is particularly a significant problem in Australian rangelands where ‘mixed’ land cover types (ie. mixture of live and died vegetation, bare soil and stony materials) are typical within the resolution of available remotely sensed data (eg. Landsat TM). For years, research efforts have been made in Australian rangelands (eg. Graetz and Gentle, 1982; Pech, et al., 1986; Graetz, et al.; 1988; Pickup, et al., 1993; and Williamson and Eldridge, 1993) and elsewhere (eg. Smith, et al., 1990; Dymond, et al., 1992; and Anderson, et al., 1993), and encouraging results were obtained. Tasks remain, however, to model sub-pixel components and their contributions to spectral measurements of satellite images, to improve the relatively low accuracy in quantifying vegetation cover or biomass, and to overcome difficulties in transferring the techniques on an operational scale. Accurate field estimation of vegetation cover and biomass is essential for any improvements in rangeland vegetation modelling and monitoring. It is highly arguable on the accuracy and correctness of some current field techniques for vegetation cover and biomass estimation (Wilson, et al., 1987) due to the common constraints of available time and fund. In addition to the sample size issues outlined by Curran and Williamson (1986), it has also been argued that some commonly used field estimation methods can produce highly subjective and inconsistent results in the ‘ground truthing’, making the subsequent image processing of remotely sensed data a meaningless effort. This paper reports a recent research project which focuses on improved field techniques for rangeland vegetation investigation. A semi-arid rangeland field station has been selected for intensive field testing of a number of field techniques employed by pasture management and remote sensing. The field sampling results from visual estimation, line intercept, quadrat and 2-D crown cover model have been compared with orthographic digital images which were acquired from a digital camera. The statistical relationships between results obtained from different methods and observers were analysed for their suitability for ground truthing of rangeland remote sensing.
2.0 METHODOLOGY The study was undertaken in Fowlers Gap Arid Zone Research Station where bladder saltbush (Atriplex vesicaria) dominated rangeland is typical for western New South Wales, Australia. A comprehensive GIS data base has been established for the Station since 1987 and subsequently updated. Image and GIS data processing were conducted using the research infrastructure supported in the School of Geography and Centre for Remote Sensing and GIS at the University of New South Wales.
2.1 FIELD INVESTIGATION The field investigation included two major tasks, namely cover and biomass measurements. For the purpose of this paper, only the first, the cover estimation, will be discussed. The differential GPS
instruments and technique were employed to accurately locate field sample sites and ground control points for subsequent image rectification with a maximum 2-D locational error of 5 metres.
2.1.1 Sampling Strategy Two transects, each contains 7 sample sites 500 metres apart, were sampled in two field trips in the summer (December 1994) and autumn (April 1995). The first transect (Transect A) is located in a paddock where a mixture of ephemeral and perennial grasses and shrubs is the dominant vegetation community with a relatively high vegetation cover. The second transect (Transect C) is located in a paddock where bladder saltbush (Atriplex vesicaria) is dominant with a low vegetation cover and virtually no grass. Each sample site covers an area of 100 × 100 metres to ensure the sample sites be overlapped with geometrically rectified TM and SPOT data. For each site, ground sampling has been undertaken at two levels, namely, site and subplot level. At the site level, two methods were used including visual estimation and line intercept. Ten 2 × 2 metre subplots were randomly located within each site to form a collection of subplots. At the subplot level, three cover sampling methods were used, namely, visual estimation, digital photography and 2-D measurements of crown cover model for individual plant.
2.1.2 Cover Measurements Vegetation cover estimation was undertaken at site and subplot levels. At the site level (100 × 100 metres), each observer was asked to visually estimate percentage vegetation cover for the site. Line intercept method was also used by applying two crossing 100 metre measuring tapes. For every 10 cm, an observation was made on if the tape intercepted with vegetation or bare ground. Percentage vegetation cover (Av) is then calculated as:
Av =
∑L
v
L
× 100
where Lv is the length of vegetated area intercepted with the tape, L is the total length of the tape (200 m). At the subplot level, individual observer was asked to visually estimate percentage vegetation cover in the 2 × 2 metre area.
(1)
b a
Each plant in the quadrat was measured by its long axis and the axis perpendicular to the long axis. The measurements are used in Figure 1. The 2-D crown cover an ellipse model to calculate the area of the plant as Ap = πab (see model: a denotes the radius of Figure 1), thus the percentage area can be aggregated for each subplot the long axis; b is that of the by adding all plant areas then dividing the sum by the total area of the axis perpendicular to a. plot (A = 4 m2):
Av =
∑A A
p
× 100
(2)
Many existing remote sensing ‘ground truthing’ techniques involve visual estimation of ground cover. This method, however, is highly subjective and largely dependent upon the experience of the field investigator. This is particularly true for Australian rangeland where the major type of vegetation cover is perennial shrubs which often create side-looking ‘false vision’ to the viewer due to their nature as scattered standing plants (Plate 1). To avoid this problem, a method which produces vertical orthographic view of the ground has been developed using a digital still-frame video camera, which has previously been reported to be suitable for field estimates of light intercepted by shrubs and low leaf area index (LAI) in open shrublands (Law, 1995). A low-cost digital camera Plate 1. The side-looking vision of the scattered standing plants. (Apple Quicktake 100) was mounted on a 5-metrehigh stand. With the camera’s 640 × 480 pixel resolution and equivalent 50-mm focus length, ground images were acquired with a 24 bit, 5 × 5 mm pixel resolution and minimum geometric distortion (Plate 2). This has ensured an accurate and objective vision of the field vegetation cover. The images were then visually interpreted and classified to create classified images containing two classes - vegetation and bare ground (Plate 3).
Plate 2. The digital vegetation image acquired.
Plate 3. Interpreted vegetation cover from Plate 2 (Dark areas are vegetation; light area is bare soil).
The visual interpretation was undertaken using ESRI ArcView 2 GIS and Corel Photo-Paint graphics software and two sets of comparable results have been independently produced by two observers. Using the ArcView software, ‘vegetation polygons’ were drawn on the screen and area were calculated and recorded manually. Using the Corel Photo-Paint software, ‘vegetation pixels’ were drawn on the screen and subsequently counted by inputting the classified images into the PCI image processing software.
2.2 STATISTICAL ANALYSIS To compare with the site-level measurements, the subplot-level data are aggregated into the site level by averaging the subplots of the site. Thus, at the site level, six sets of 14 samples are produced, namely, site visual estimates, averaging subplot visual estimates of four observers, line intercept, averaging subplot 2-D crown cover models, and two sets of averaging subplot image interpretation results. At the subplot level, four sets of 140 samples are produced, namely, averaging visual estimates of all observers, 2-D crown cover models, and two sets of image interpretation results. In addition to these, visual estimates made by individual observer (four sets for all subplots) are also available. Statistical analysis at both site and subplot levels has been carried out using Microsoft Excel data analysis functions. Correlations between each field method has been computed and results at site and subplot levels have been derived for the discussion.
3.0 RESULTS AND DISCUSSION The correlations between different methods at the subplot level are summarised in Table 1 which shows that all methods performed at the subplot are significantly correlated with 2-D crown cover model producing least correlated results. The weaker correlation between the 2-D crown cover model and other methods is expected and can be explained by the fact that small plants in the subplot were not sampled due to practical difficulties. Table 1. Correlation coefficients between different methods and observers at subplot level (all transects). II2 II2 II1 CCM APV Ob1 Ob2 Ob3 Ob4
1 0.9194 0.8905 0.9089 0.9034 0.8227 0.8771 0.7666
II1 1 0.8589 0.9174 0.9130 0.8330 0.9033 0.7971
CCM
1 0.8621 0.8469 0.7773 0.8566 0.7368
APV
1 0.9536 0.9334 0.9655 0.9117
Ob1
Ob2
Ob3
1 0.8405 0.9325 0.8146
1 0.8829 0.8038
1 0.8415
Ob4
1
where: APV = Average subplot visual estimates; CCM = 2-D crown cover model; II1 = Image interpretation observer 1; II2 = Image interpretation observer 2; Ob1 = Observer 1, Ob2 = Observer 2; Ob3 = Observer 3; Ob4 = Observer 4. * All significant at 0.001 level n = 140
Further examination on individual transect has unveiled that, though still with a high significance level, correlations between the methods become much weaker while investigating land with less vegetation cover, such as Transect C (Table 2). All methods tend to preform better while dealing with lands with higher vegetation cover, but yield more inconsistent results with a reduced cover level.
Table 2. Correlation coefficients between different methods and observers at subplot level (Transect C). II2 II2 II1 CCM APV Ob1 Ob2 Ob3 Ob4
II1
1 0.9103 0.7367 0.8925 0.8690 0.7698 0.8159 0.7518
CCM
1 0.7975 0.8649 0.8322 0.7943 0.8047 0.7038
APV
1 0.8620 0.7735 0.7386 0.8275 0.7620
1 0.9036 0.8685 0.9246 0.8968
Ob1
Ob2
Ob3
Ob4
1 0.7299 0.8360 0.6682
1 0.7418 0.7490
1 0.7584
1
where: APV = Average subplot visual estimates; CCM = 2-D crown cover model; II1 = Image interpretation observer 1; II2 = Image interpretation observer 2; Ob1 = Observer 1, Ob2 = Observer 2; Ob3 = Observer 3; Ob4 = Observer 4. * All significant at 0.001 level n = 70
Correlations between visual estimates by different observers are also significantly correlated at the subplot level (Table 1 and Table 2) though the trend is also shown that less consistent results are observed while the cover level is reduced. It is perhaps more important to show that the correlations between observers are weaker than those between different methods, compared in both Table 1 and Table 2, demonstrating the weakness and inconsistency of the visual estimation method. More controversial results have been obtained at site level where a distinguished inconsistency has been shown between the methods of direct measurements at the site level (average site visual estimates and line intercept) and the methods using aggregated data from subplots (Table 3). The line intercept method has produced weakest correlation with all other methods while averaged site visual estimates has produced weak yet acceptable correlations with others. Table 3. Correlation coefficients between different methods at site level (all transects). ASV!
APV
LI
CCM
II1
II2
ASV
1
APV
0.9389*
1
0.6442**
0.7657*
1
CCM
0.8497*
0.9215*
0.6556**
1
II1.
0.8398*
0.9138*
0.7957*
0.8882*
1
II2
0.8446*
0.9175*
0.6736**
0.9469*
0.9411*
LI
1
!
ASV = Average site visual estimates; APV = Average subplot visual estimates; LI = Line intercept; CCM = 2-D crown cover model; II1 = Image interpretation observer 1; II2 = Image interpretation observer 2. * Significant at 0.001 level ** Significant at 0.01 level n = 14
With poor and scattered vegetation cover types (eg. Transect C), the weakness of visual estimation has clearly been shown as it yields virtually no correlation to any other methods (Table 4). The trend shown at subplot level has also been confirmed here that inconsistent results tend to be associated with reduced vegetation cover. It has been demonstrated that this trend is further strengthened with data aggregation errors by comparing Table 3 and Table 4. It is, however, worth to notice that the correlation between two observers’ image interpretation results has been significantly consistent through all cases.
Table 4. Correlation coefficients between different methods at site level (Transect C). ASV!
APV
LI
CCM
II1
II2
ASV
1
APV
0.3337
1
-0.0615
0.5974
1
CCM
0.0798
0.7449***
0.3957
1
II1.
0.0930
0.7226***
0.3030
0.3656
1
II2
0.3570
0.8625**
0.4233
0.3970
0.8750**
LI
1
!
ASV = Average site visual estimates; APV = Average subplot visual estimates; LI = Line intercept; CCM = 2-D crown cover model; II1 = Image interpretation observer 1; II2 = Image interpretation observer 2. * Significant at 0.001 level ** Significant at 0.01 level *** Significant at 0.05 level n=7
The inconsistency between observers’ visual estimation results has been significant at the site level (Table 5), with generally weaker correlation between the observers than that between different methods (comparing Table 3). Table 5. Correlation coefficients between observer’s site visual estimates (all transects). Observer 1
Observer 2
Observer 3
Observer 1
1
Observer 2
0.6763**
1
Observer 3
0.8729*
0.9351*
1
Observer 4
0.7406**
0.9421*
0.8863*
* Significant at 0.001 level
** Significant at 0.01 level
Observer 4
1
n = 14
The following points can therefore be concluded by this study: a) All methods tested are significantly correlated at the subplot level where no data aggregation is involved. b) The correlations become weaker at the site level where data aggregation errors have made the results much less consistent. It can be argued that data aggregation can introduce much higher levels of error than sampling methods themselves. c) The lower consistency has been observed when vegetation cover become poorer with scattered standing vegetation types. d) A Higher error level has been observed between observers’ visual estimates than that between different methods, showing that visual estimates have produced the least reliable results.
e) High-level of consistency has been observed between image interpretation results by different observers, demonstrating that the photographic method has a significant potential to be an objective and highly accurate field investigation method.
4.0 CONCLUSION The statistical analysis of the field sampling results has shown that a high-level of errors has been associated with data aggregation while conducting rangeland vegetation ground truthing. The commonly used visual estimation method in rangeland remote sensing has delivered most inconsistent and unreliable results. A new method has been invented using a remotely controlled digital camera with associated image interpretation methods which allow accurate and consistent estimates of vegetation cover. This method can be further enhanced with image classification techniques to enable accurate, automated and objective ground truthing for rangeland remote sensing. Based on these detailed investigations, arguments have been presented on the accuracy and correctness of commonly-used field investigation methods. The results have shown that care must be taken on the use of some of these methods for rangeland vegetation cover and biomass estimation, which can in turn largely influence the accuracy of image processing for vegetation modelling based on the ground truthing.
5.0 ACKNOWLEDGEMENT The project was funded by Australian Research Council. The authors would like to thank field investigation team members Allan Evans, Richard Jossop, Ben McDonald, Tim Loarn, Marc Robson and John Cuff for their assistance and contributions to the field data collection. 6.0 REFERENCES Anderson, G.L., Hanson, J.D. and Haas, R.H., “Evaluating Landsat Thematic Mapper derived vegetation indices for estimating above-ground biomass on semiarid rangelands”, Remote Sensing of Environment, Vol. 45, No. 2, pp. 165-175, 1993. Curran, P.J. and Williamson, H.D., “Sample size for ground and remotely sensed data”, Remote Sensing of Environment, Vol. 20, No. 1, pp. 31-41, 1986. Dymond, J.R., Stephens, P.R., Newsome, P.F. and Wilde, R.H., “Percentatge vegetation cover of a degrading rangeland from SPOT”, International Journal of Remote Sensing, Vol. 13, No. 11, pp. 1999-2007, 1992. Frazier, P.S., Applegate, R.J., Wood, B.G. and Hill, G.J.E., “The role of satellite remote sensing in rangeland monitoring: a report from the national rangeland monitoring program workshop”, In Proceedings of 7th Australasian Remote Sensing Conference, Melbourne, pp. 1148-1155, 1-4 March 1994. Graetz, R.D. and Gentle, M.R., “The relationships between reflectance in the Landsat wavebands and the composition of an Australian semi-arid shrub rangeland”, Photogrammetric Engineering and Remote Sensing, Vol. 48, No. 11, pp. 1721-1730, 1982.
Graetz, R.D., Pech, R.P. and Davis, A.W., “The assessment and monitoring of sparsely vegetated rangelands using calibrated Landsat data”, International Journal of Remote Sensing, Vol. 9, No. 7, pp. 1201-1222, 1988. Law, B.E., “Estimation of leaf area index and light intercepted by shrubs from digital videography”, Remote Sensing of Environment, Vol. 51, No. 2, pp. 276-280, 1995. Pech, R.P., Graetz, R.D. and Davis, A.W., “Reflectance modelling and derivation of vegetation indices for an Australian semi-arid shrubland”, International Journal of Remote Sensing, Vol. 7, No. 3, pp. 389-403, 1986. Pickup, G., Chewings, V.H. and Nelson, D.J., “Estimating changes in vegetation cover over time in arid rangelands using Landsat MSS data”, Remote Sensing of Environment, Vol. 43, No. 3, pp. 243263, 1993. Smith, M.O., Ustin, S.L., Adams, J.B. and Gillespie, A.R., “Vegetation in deserts: I. A regional measure of abundance from multispectral images”, Remote Sensing of Environment, Vol. 31, No. 1, pp. 126, 1990. Williamson, H.D. and Eldridge, D.J., “Pasture status in a semi-arid grassland”, International Journal of Remote Sensing, Vol. 14, No. 13, pp. 2535-2546, 1993. Wilson, A.D., Abraham, N.A., Barratt, R., Choate, J., Green, D.R., Harland, R.J., Oxley, R.E. and Stanley, R.J., “Evaluation of methods of assessing vegetation change in the semi-arid rangelands of southern Australia”, Australian Rangeland Journal, Vol. 9, No. 1, pp. 5-13, 1987.