|Pu, Ruiliang - UC BERKELEY|
|Gong, Peng - UC BERKELEY|
|Yong, Tian - UC BERKELEY|
|Xin, Miao - UC BERKELEY|
Submitted to: Environmental Monitoring and Assessment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: May 18, 2007
Publication Date: June 28, 2007
Citation: Pu, R., Gong, P., Yong, T., Xin, M., Carruthers, R.I., Anderson, G.L. 2007. Invasive species change detection using artificial neural networks and CASI hyperspectral imagery. International Journal of Remote Sensing. 140:15-32. Interpretive Summary: Remote sensing was used to assess the control effectiveness of a new USDA-ARS developed method of invasive plant control. A biological control agent, a leaf-feeding beetle from China, was introduced into the western US following several years of safety testing and impact evaluation that was conducted under US quarantine conditions. Once released, USDA and cooperators assess the impact of such biological control agent using extensive ground sampling. This particular insect natural enemy, however, was so effective that it impacted wide areas within a very short period of time. Remote sensing was thus investigated as a cost effective tool to determine the extent to which the introduced biological control agent defoliated the invasive weed, saltcedar, and to the extent that it spread across the test areas. It was found that hyperspectral remote sensing provided an excellent measure of defoliation and showed the extent that the USDA program was effectively working. This is important as it provides USDA and cooperators with a new cost effective tool to help in wider assessment that will need to be completed beginning in the spring and summer of 2006. This is when USDA is beginning a new saltcedar biological implementation program where these beneficial agents will be spread widely in 14 western states.
Technical Abstract: A change detection experiment for an invasive species, saltcedar, near Lovelock, Nevada, was conducted with multitemporal Compact Airborne Spectrographic Imager (CASI) hyperspectral data sets. The data were acquired on July 2, and August 29, 2002, and September 10, 2003. Post-classification and pre-classification change detection methods were tested. In the post-classification strategy, a principal component analysis (PCA) was separately performed to single-date CASI imagery in the visible bands and NIR bands. Then the first 5 PCs from the visible bands and the first 5 from the NIR bands were used to classify 6-8 cover types with a maximum likelihood classifier. A complete matrix of change information was produced by overlaying two single-date classification maps. In the pre-classification strategy, a linear regression model was simulated between two normalized difference vegetation index (NDVI) images to normalize spectral differences caused by factors not related to land cover change. Then the actual time 2 NDVI image was subtracted by the predicted time 2 NDVI image to obtain the differencing image. The NDVI differencing image was further processed with thresholds into change/no-change of salt cedar. By testing the single-date classification results and validating the change/no-change results, both change detection results have indicated that CASI hyperspectral data had potential in mapping and monitoring the change of saltcedar. Accuracy assessment and change/no-change validation for both methods are presented. According to the criteria of quickly filtering out changed pixels from non-changed background and using fewer spectral bands, the NDVI differencing (pre-classification) method is recommended if a suitable spectral normalization can be done before performing image differencing.