Forest Monitoring using all Archived Landsat Data
Thonfeld, Frank¹; Menz, Gunter^2
1Center for Remote Sensing of Land Surfaces (ZFL), University of Bonn, GERMANY; ²Remote Sensing Research Group (RSRG), University of Bonn, GERMANY

Forest monitoring is essential for the assessment of CO2 exchange rates, biomass estimation, forest health assessment and the observation of environmental conditions. Remote sensing allows for the monitoring of spacious areas supporting agencies in the field. For some decades it was state of the art to classify images and apply post-classification comparison to assess changes. Using those bi-temporal approaches allows for the detection of apparent forest conversion, i.e. the replacement of one land cover by another through logging or fire. However, seasonal fluctuations and long-term trends cannot be captured with bi-temporal datasets and the above mentioned methods. Over the past years several methods were developed to explore annual Landsat images. The images for resulting trajectories have to be selected very carefully since sun angle differences, sun-sensor-target geometry, phenology, water availability and other environmental conditions affect strongly the comparability of images. Using anniversary data allows for the assessment of abrupt changes as well as long-term trends. Changes between consecutive images such as seasonal or interannual variations cannot be precisely assessed. With the availability of standard Modis products it is possible to apply time series analysis. They allow for the detection of seasonal patterns, abrupt changes and long term trends. The disadvantage of Modis products is their coarse spatial resolution compared to Landsat. Unfortunately, it is not possible to create regular Landsat products comparable to the Modis 16-day NDVI product (or others). Favored by the opening of the Landsat archive in 2008, however, it is possible to combine the advantages of dense time series (i.e. detection of seasonal patterns, abrupt changes, long-term trends) and high spatial resolution. Here, we make use of the bfastmonitor method to detect abrupt changes in forests by analyzing all available Landsat scenes over the southern part of Vancouver Island, Canada. All Landsat 5 and 7 images of path/row 48/26 and 47/26 were processed to create a layerstack of 1014 scenes ranging from January 2000 to December 2012. The data were atmospherically corrected using NASA’s LEDAPS software. Since cloud and cloud shadow masking is essential for further analysis, FMASK was applied. Relative radiometric normalization was conducted by means of Iteratively Re-Weighted Multivariate Alteration Detection (IR-Mad) to harmonize the images in the time series. For each image several vegetation indices were calculated. Here, we refer only to the NDVI since it is considered appropriate for the purpose of forest monitoring. The main challenge is that the Landsat NDVI time series is not regular due to cloud cover, SLC-Off stripes, missing data etc. Thus, after applying the FMASK cloud mask each pixel has its own individual time series with gaps at different positions. The interval between scenes varies between 1 and 9 days with one gap being 16 days. From the theoretically possible 1014 observations in the overlapping part of the two tiles up to 390 observations show clear land surface. The preliminary results presented here are restricted to that overlapping area. Based on bfastmonitor, the timing and magnitude of all abrupt changes were detected and mapped. The interpretation of numerous samples revealed that true changes were well detected in space and time irrespective of the number of clear observations. However, in the areas with no true changes small magnitude spurious changes were sometimes detected due to noisy time series.
It can be shown that Landsat results show similar patterns than Modis results with Landsat providing much more spatial details. Thus, forest monitoring with all available Landsat information allows for the detection of abrupt changes with high temporal (comparable with Modis time series) and spatial accuracy (more details than Modis).