UC Berkeley

We developed a semi-automatic algorithm named Berkeley Indices Trajectory Extractor (BITE) to detect forest disturbances, especially slow-onset disturbances such as insect mortality, from time series of Landsat 5 Thematic Mapper (TM) images. BITE is a streamlined process that features trajectory extraction and interpretation of multiple spectral indices followed by an integration of all indices. The algorithm was tested over Grand County in Colorado, located in the Southern Rocky Mountains Ecoregion, where forests dominated by lodgepole pine have been under mountain pine beetle attack since 2000. We produced a disturbance map using BITE with an identification accuracy of 94.7% assessed from 602 validation sample pixels. The algorithm shows its robustness in deriving forest disturbance type and timing with the presence of different levels of atmospheric conditions, noises, pixel misregistration and residual cloud/snow cover in the imagery. Outputs of the BITE algorithm could be used in studies designed to increase understanding of the mechanisms of mountain pine beetle dispersal and tree mortality, as well as other types of forest disturbances.

Large remote sensing datasets, that either cover large areas or have high spatial resolution, are often a burden for information mining for scientific studies. Here, we present an approach that conducts clustering after gray-level vector reduction. In this manner, the speed of clustering can be considerably improved. The approach features applying eigenspace transformation to the dataset followed by compressing the data in the eigenspace and storing them in coded matrices and vectors. The clustering process takes advantage of the reduced size of the compressed data and thus reduces computational complexity. We name this approach Clustering Based on Eigen Space Transformation (CBEST). In our experiment with a subscene of Landsat Thematic Mapper (TM) imagery, CBEST was found to be able to improve speed considerably over conventional K-means as the volume of data to be clustered increases. We assessed information loss and several other factors. In addition, we evaluated the effectiveness of CBEST in mapping land cover/use with the same image that was acquired over Guangzhou City, South China and an AVIRIS hyperspectral image over Cappocanoe County, Indiana. Using reference data we assessed the accuracies for both CBEST and conventional K-means and we found that the CBEST was not negatively affected by information loss during compression in practice. We then applied CBEST in mapping the forest change from 1986-2011 for the entire state of California, USA with over 400 Landsat TM images. We discussed potential applications of the fast clustering algorithm in dealing with large datasets in remote sensing studies.

We present an efficient approach for a practice of large-area mapping of forest changes based on the Clustering Based on Eigen Space Transformation (CBEST) algorithm using remote sensing. By analyzing 450 Landsat Thematic Mapper (TM) satellite images from 1986 to 2011 with a five-year interval covering the entire state of California, USA, we derived a forest change type map, a forest loss map and a forest gain map. Although California has 99.6 million acres land area in total and the spatial resolution of Landsat TM is 30m, the computing time of the task took only 10 hours in a computer with an Intel 2.8 Ghz i5 CPU and 8 Gigabytes RAM. The overall accuracy of the forest cover in year 2011 was reported as 92.9% ± 1.6%. We found that the estimated forest area changed from 28.20 ± 1.98 million acres to 28.05 ± 1.98 million acres from 1986-2011. In particular, our rough estimate indicates that each year California's forest experienced loss of 92 thousand acres and recovery of 85 thousand acres, resulting in seven thousand acres forest loss per year. In addition, during 1986-2011, around 12% of the forestland experienced changes, in which the change was 4% each for deforestation, afforestation and deforestation then recovered respectively. We concluded that the forestland in California had been managed in a sustainable manner over the 25 years, since no significantly directional changes were observed. Our approach made a tighter estimate of the true canopy coverage such that 29% of land in California is forestland, comparing with the statistics of 33% and 40% made by previous studies that had lower spatial resolution and shorter temporal coverage.