Open smithve87 opened 7 years ago
bendv
commented
7 years ago Why are you only using imagery from 2005 onwards? How is your history period defined? High seasonality in tropical dry forests is known to create complications in bfastmonitor. Check out the following for possible strategies to deal with this:
http://dx.doi.org/10.1016/j.isprsjprs.2015.03.015 http://dx.doi.org/10.1016/j.rse.2015.11.006
smithve87
commented
7 years ago Thank you for the response. I have read the articles and made some adjustments. I have gone back and used data dating back to 1985 now in the hopes of adding more so the seasonal trend could be more well defined. As for the history period, it is from 1985 - 2014, with the monitoring period being the full year of 2015. I have also changed many of the parameters such as the h value (0.25 or 0.5 produced results), harmonic order (1 - 3), model (trend + harmon, harmon only, trend only), and in various combinations. However, the breakpoints that are detected still aren't quite matching up. Next I will try using only Landsat images with low cloud cover (10% - 20% max scene cover), as well as trying to use MODIS data as well. Would you suggest any other strategies to reduce noise and improve breakpoint accuracy?
I am attempting to track deforestation in tropical dry forests in the Yucatan in Mexico and Guanacaste in Costa Rica. Tropical dry forests present unique challenges in that they have at least one pronounced dry season during the year where the deciduous vegetation experiences senescence. I am using Landsat scenes from 2005 until present that were preprocessed in the USGA ESPA ordering system. Poor quality and high cloud cover scenes were not included in the scene list, and ESPA preprocessing consisted of CFMask application, extent modification and inclusion of vegetation indices (NDVI, EVI, NBR, NBR2, MSAVI, NDMI). I've successfully run bfmSpatial with varying parameters such as changing monitoring period length (3 months, 6 months, 1 year) as well as tweaking parameters such as the formula type (trend, harmon, both), h value, order, end, level, etc. However, when comparing the results to high-res imagery, it seems that bfmSpatial isn't particularly accurate. My principal investigator, who is highly experienced in classifying deforestation in imagery, especially high-res images, created the fishnet or grid and identified locations of deforestation, potential deforestation and degradation. When comparing the bfmSpatial results against the areas of deforestation, there is not much crossover so I'm wondering if others have had similar issues. Perhaps I need more historical data; to include all images regardless of quality or cloud cover; tweak different parameters; etc.? Or maybe it's just that the nature of a tropical dry forest with pronounced seasonality could be misinterpreted as deforestation. I'm wondering if anyone has had experience in regard to detection accuracy, especially in studying tropical dry forests. Thank you for any and all suggestions, and I can provided any needed files.