COMPARISON OF CHANGE DETECTION TECHNIQUES FOR THE YUCATAN PENINSULA USING LANDSAT IMAGE TIME SERIES René R. Colditz, Ricardo M. Llamas, Steffen Gebhardt, Thilo Wehrmann, Julian Equihua National Commission for the Knowledge and Use of Biodiversity (CONABIO), Avenida Liga Periférico – Insurgentes Sur No 4903, Col. Parques del Pedregal, Del. Tlalpan 14010 Mexico, D.F. Tel: +52-55-50045020, email:
[email protected] ABSTRACT Several change detection methods have been developed over the last decades and an even higher number during the last years due to the opening of the Landsat archive. Some change detection methods aim at all types of land surface alterations while others target specific types such as inundations, urbanization or forest cover change or even more specifically forest cover loss. Many methods were developed and tested for temperate regions where most cloud-free data are obtained during the growing season. In the tropics, however, cloud cover is highest during the period when vegetation is most active. This study tests two common approaches, the Vegetation Change Tracker (VCT) and the Iterative Multivariate Alteration Detection (IMAD) in the northwestern portion of the Yucatan peninsula using Landsat images. Results from change detection algorithm were compared to reference samples and reference polygons. Various parameter sets for the VCT algorithm never reach the accuracy level of IMAD. Index Terms— Change detection, Vegetation Change Tracker (VCT), Iterative Multivariate Alteration Detection (IMAD)Yucatan, MODIS 1. INTRODUCTION During the last four decades multiple techniques have been developed for change detection using satellite images [1, 2]. The success of any technique depends on various aspects, such as available resources, time for completing the study, image availability, accessibility to ground observations and ancillary data, availability and experience with change detection algorithms, area of expertise, intended use of the product, etc. [3]. Equally important is the type or kind of change to be identified, i.e. a conversion or a modification of land cover [1] and the speed at which the change occurs, i.e. an abrupt event or gradual transformation as a trend over several years. In recent years, specific techniques have been developed for each change type and process, many with a particular emphasis on forest cover change at different spatial scales [4-6].
978-1-4799-7929-5/15/$31.00 ©2015 IEEE
Most methods were tested in temperate regions where vegetation growth is linked to the warmer temperature during spring and summer seasons when there is also an increased availability of cloud-free satellite observations. In outer tropical areas vegetation growth is connected to the amount of available water in form of precipitation; in the inner tropics the availability of light seems to be an additional driver [7]. Therefore, in tropical areas the peak of vegetation growth at the end of the rainy season complicates change detection using optical satellite data due to persistent cloud cover. The study area of the Yucatan peninsula in the outer tropics, specifically Landsat path-row 020-046, is a challenging region as there are several types of anthropogenic (shifting cultivation, pasture land use, urban expansion), and natural (hurricanes, fires, coastline modifications) changes. The selected area forms a test site for forest carbon modeling for which the results of this study will be used as activity data. Recently developed national products of annual land cover change (2005-2011) based on MODIS images have revealed some changes but the automated change detection technique is limited by the coarse spatial resolution of 250m pixel size. [8]. This study aims at exploring two techniques, the Vegetation Change Tracker (VCT) and the Iterative Multivariate Alteration Detection (IMAD) algorithm, for identifying changes that occur in the Yucatan peninsula using a time series of historical and recent 30m Landsat data. 2. DATA AND METHODS The Landsat 5TM and 7ETM+ data archive of path-row 020046, which corresponds to the northwestern sector of the Yucatan peninsula South of Merida, was downloaded from the United States Geological Survey (USGS). For change detection using the Vegetation Change Tracker (VCT) [5], the data archive was screened and a set of cloud-free images was extracted. For the entire period from 1985-2010 dryseason images mostly between December and May were selected. Landsat 7ETM+ data were used for the period 2001 to 2009 due to the discontinuation of Landsat 5TM over
1650
IGARSS 2015
southern Mexico. From 2003 onwards, two Landsat 7ETM+ images were mosaicked to reduce data gaps due to scan-line errors and cloud coverage. Due to specific requirements by the VCT algorithm images were preprocessed by Earth Science Processing Architecture (ESPA) on-demand interface of the USGS, which also provides additional masks for clouds and cloud shadow from which we used Fmask [9]. There are two important parameters to be tested in this study the Normalized Difference Vegetation Index (NDVI) and the Vegetation Continuous Fields (VCF) product, which both define the spatial extent of mature forest. The Iterative Multivariate Alteration Detection (IMAD) algorithm was applied to atmospherically corrected and cloud-screened Landsat data using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) [10] and Fmask algorithms [9]. The multivariate alteration detection (MAD) algorithm is a bi-temporal change detection technique which acts upon two N-band images of the same scene acquired at different times. Canonical correlation analysis is performed to produce linear combinations of each of the two sets of image bands that show maximum correlation between each other. There are up to N of these combinations (canonical correlation components), these are ordered by their mutual correlation and are pair-wise orthogonal between each other. In addition they are invariant under affine transformations of the input images which makes this procedure resistant to linear effects introduced by atmospheric conditions and sensor calibrations [11]. The subtraction of these canonical correlation components are referred to as MAD components and can be used as a convenient background for change detection. The sensitivity of the MAD transformation can be augmented by weighing the observations (pixels) in an iterative manner by their probability of no-change in the preceding iteration. This iterative MAD algorithm is referred to as the IMAD transformation. The spatial congruence of the difference signal in the IMAD components may be further improved by subsequently applying a Maximum Autocorrelation Factor (MAF) transformation [12]. This finds linear combinations of the IMAD components that show maximum spatial autocorrelation measured by comparing the IMAD image with itself shifted one pixel right-upwards. Change products were compared to manually digitized polygons for intervals 1995-2000, 2000-2005, 2005-2010, respectively, which served as reference data for accuracy assessment. This assessment focuses on areal agreement and disagreement (commission and omission error) against the pixels classified as change with VCT and IMAD. A second set uses 300 sample points and follows the classical way of a random stratified sample design. Samples were drawn the following way: 100 samples within change areas as defined in the polygon reference set, 100 samples in a buffer zone of 90m outside the change polygons and 100 samples in the rest of the image.
Table 1. Properties of reference polygons. Properties
1995-2000
2000-2005
2005-2010
Initial Date
March 22
April 20
April 10, April 26
Final date
April 20 2
Total study area [km ] Number of polygons Total change area [ha] Min. change area [ha] Max. change area [ha] Mean change area [ha]
31,712 7,892 106,733 5 872 13.52
April 10, April 26 33,910 5,063 62,454 5 4,371 12.33
January 26 33,711 2,804 40,048 5 731 14.28
3. RESULTS Table 1 presents the results for reference polygons. A high difference is noted in the number of polygons and the area estimated as change for the period 1995-2000 in comparison to the others, which impacts the assessment of changes. Change results from VCT with various parameters and IMAD were assessed with reference polygons and reference sample points for three periods (Figure 1). Both assessment approaches show high errors and notable differences among the periods. Polygon-based assessment between 1995 and 2000 indicates variability among VCT parameters for the commission error but always high omission errors; on the contrary commission error is always high but omission error varies for the other assessment periods (2000-2005 and 2005-2010). Inverse patterns are depicted for the samplebased assessment. In general IMAD shows better results than VCT. Parameter sets of VCT using different thresholds for NDVI and VCF result in varying accuracies. In general, a NDVI threshold of around 0.7 can be recommended for this specific study site and selection of Landsat images. One cannot draw a conclusive pattern for the most appropriate VCF threshold. Results in Figure 2 show a generally better spatial coincidence of IMAD with reference polygons. Results of VCT for forested areas (A) appear to overestimate changes, and in suburban settings (B) the algorithm does not detect changes as the area was not considered as mature forest. This was expected, as VCT only focuses on forest changes, while IMAD is a generic change detection algorithm. The examples demonstrate nicely that IMAD in many cases resembles the visually delineated changes by the analyst but often not fully matches the shape or area 4. DISCUSSION AND CONCLUSIONS In general, all assessments show inferior than expected results for both, VCT and IMAD. Lower accuracies (higher errors) were expected for the reference set using polygons due to differences in spatial delineation of change areas. The
1651
Figure 1. Omission-commission error plots in percent for polygon and sample point-based assessment.
Figure 2. Results of IMAD and VCT compared to reference polygons for the period 2000-2005. The two subsets correspond to a forested region in Campeche state (A) and the area around the city of Merida (B).
1652
difficulties of defining and mapping change spatially by visual interpretation is inherent in our analysis. Comparing the number of polygons and the area detected in reference set 1995-2000 is much higher than for periods 2000-2005 and 2005-2010. One may expect some differences due to changes in land use over a 15-year period, but not at this scale. Therefore there is an inherent bias in our spatial reference data. We intended to compensate this bias with a random stratified sample design drawing sample points. The results clearly differ from our areal assessment but in particular omission errors are still high. This indicates that the detected changes are a conservative estimate in comparison to this reference data set. In this study we presented results for a minimum mapping unit (MMU) of 5ha, i.e. in both, reference and change detection results only polygons with an area of at least 5ha were considered. We also conducted assessments with 1ha and 10ha MMU. In most cases the results were not different and we could not find a clear pattern of improvement in either using smaller or larger minimum object sizes. Only for VCT runs with very relaxed settings for VCF and NDVI, i.e. defining mature forest broadly, a notable difference was found as higher MMU eliminated many small patches which reduced the commission error. In terms of algorithm comparison, VCT was extensively tested with various parameter sets. Further assessments (not shown) also employed reduced sets of image data to eliminate potentially remaining cloudy or hazy pixels. Still, the results of VCT almost always show lower accuracies than for IMAD. A potential explanation is the selection of images during the dry season, which was justified by more cloud-free observations. However, for VCT peak-of-growing season images are recommended [5] due to the highest spectral contrast. An additional limitation of VCT is transferability as the quality of results highly depend on the hand-selected image set and user-experience for appropriate parameters. In contrast, the IMAD algorithm is fully automated allowing for scene-based large-area processing. In follow-up studies we will consolidate and confirm our results to provide clear recommendations for carefully selecting algorithms and their performance parameters in tropical settings. Also further comparisons to, for instance, global products [13] will be included. 5. REFERENCES [1] P. Coppin, I. Jonckheere, K. Nackaerts, B. Muys, and E. Lambin, “Review ArticleDigital change detection methods in ecosystem monitoring: A review,” International Journal of Remote Sensing, 25 (9), 1565–1596, May 2004.
[3] R.E. Kennedy, P.A. Townsend, J.E. Gross, W.B. Cohen, P. Bolstad, Y.Q. Wang, and P. Adams, “Remote sensing change detection tools for natural resource managers: Understanding concepts and tradeoffs in the design of landscape monitoring projects,” Remote Sensing of Environment, 113 (7), 1382–1396, Jul. 2009. [4] D. Pouliot, R. Latifovic, N. Zabcic, L. Guindon, and I. Olthof, “Development and assessment of a 250 m spatial resolution MODIS annual land cover time series (2000–2011) for the forest region of Canada derived from change-based updating,” Remote Sensing of Environment, 140, 731-743, Jan 2014. [5] C. Huang, S.N. Goward, J.G. Masek, N. Thomas, Z. Zhu, J.E. Vogelmann, “An automated approach for reconstructing recent forest disturbance history using dense Landsat time series stacks,” Remote Sensing of Environment, 114, 183-198, 2010. [6] Z. Zhu and C.E. Woodcock, “Continuous change detection and classification of land cover using all available Landsat data,” Remote Sensing of Environment, 144, 152-171, 2014. [7] R.R. Nemani, C.D. Keeling, H. Hashimoto, W.M. Jolly, S.C. Piper, C.J. Tucker, R.B. Myneni, S.W. Running, “Climate-driven increases in global terrestrial net primary production from 1982 to 1999” Science, 300, 1560-1563, June 2003. [8] R.R. Colditz, R.M. Llamas, R.A. Ressl, “Annual land cover monitoring using 250m MODIS data for Mexico” In: IEEE International Geoscience and Remote Sensing Symposium 2014, IGARSS 2014, July 13th – 18th 2014, Québec City, Canada, 46644667, 2014. [9] Z. Zhu, and C.E. Woodcock, “Object-based cloud and cloud shadow detection in Landsat imagery,” Remote Sensing of Environment, 118, 83 – 94, 2012. [10] J.G. Masek, E.F. Vermote, N.E. Saleous, R. Wolfe, F.G. Hall, K.F. Huemmrich, F.Gao, J. Kutler, and T.-K. Lim, “A Landsat surface reflectance data set for North America, 1990-2000”, IEEE Geoscience and Remote Sensing Letters, 3 (1), 68-72, Jan. 2006. [11] A. Nielsen, K. Conradsen, J. Simpson. “Multivariate alteration detection (MAD) and MAF postprocessing in multispectral bi-temporal image data: new approaches to change detection studies,”, Remote Sensing of Environment, 64, 1–19, April 1998. [12] A. Nielsen. “The Regularized Iteratively Reweighted MAD Method for Change Detection in Multi- and Hyperspectral Data,” IEEE Transactions on Image Processing, 16, 463–478, Feb 2007. [13] M.C. Hansen, P.V. Potapov, R. Moore, M. Hancher, S.A. Turubanova, A. Tyukavina, D. Thau, S.V. Stehman, S.J. Goetz, T.R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C.O. Justice, J.R.G. Townshend, “High-resolution global maps of 21stcentury forest cover change” Science, 342, 850-853, Nov 2013.
[2] D. Lu, P. Mausel, E. Brondízio, and E. Moran, “Change detection techniques,” International Journal of Remote Sensing, 25 (12), 2365–2401, Jun. 2004.
1653
