Disturbance type identification in eastern forests

Background/Question/Methods

Large scale climate models require spatially-explicit estimates of carbon flux from forest disturbance. As part a larger project to detect and annotate North American forest disturbances, we evaluated the ability of the Thematic Mapper sensors aboard the Landsat 4, 5, and 7 satellites to discriminate between fire, logging, and insect damage in Eastern Forests. Landsat TM imagery from 1984 to 2011 was compiled as predictor data. Disturbance and disease information from the US Forest Service Aerial Detection Surveys (ADS) geodatabase and disturbed plots from the US Forest Service Forest Inventory and Analysis (FIA) database provided training data for a neural network tasked with discriminating between disturbance types. Inputs to the network included Tassel-Cap transformed spectral bands, the spatial-variance of these bands over a 3x3 window, and the deviation of band-6 from the spatially- and temporally-local mean. To capture phenology, these values were computed for both a late spring and late summer interval. Additionally, to incorporate temporal change, we included inputs calculated for two years before and after the year of interest. We tested incrementally complex networks, with various combinations of the above inputs. Preliminary validation was performed using a reserved test dataset from the ADS and FIA data.

Results/Conclusions

Although there have been recent advances in discriminating between disturbance types in boreal forests, due to the larger number of forest species and cosmopolitan nature of overstory communities, separation remains difficult in eastern forests. Fire and clear cutting can be identified without phenological information (~15% misclassification); temporal information, however, improves accuracy by separating fire scars from periurban and agricultural land (~4%). Phenological and temporal information was necessary for discrimination of insect disturbances. In the full model with 50 hidden nodes, misclassification error of test data was approximately 20%, with most error caused by classifying insect disturbance as unharmed forest (~50%). The ADS database, derived from sketch maps and later digitized, commonly designates a single large area encompassing many smaller effected areas as disturbed, overestimating disturbance and creating ambiguity in the neural network. Combining this information with independently-created disturbance intensity maps during training may lead to better classification.