Joshua J. Picotte

Validation of remote sensing of burn severity in south-eastern US ecosystems

We assessed an existing method of remote sensing of wildland fire burn severity for its applicability in south-eastern USA vegetation types. This method uses Landsat satellite imagery to calculate the Normalised Burn Ratio (NBR) of reflectance bands sensitive to fire effects, and the change in NBR from pre- to post fire (dNBR) to estimate burn severity. To ground-truth ranges of NBR and dNBR that correspond to levels of burn severity, we measured severity using the Composite Burn Index at 731 locations stratified by plant community type, season of measurement, and time since fire. Best-fit curves relating Composite Burn Index to NBR or dNBR were used to determine reflectance value breakpoints that delimit levels of burn severity. Remotely estimated levels of burn severity within 3 months following fire had an average of 78% agreement with ground measurements using NBR and 75% agreement using dNBR. However, percentage agreement varied among habitat types and season of measurement, with either NBR or dNBR being advantageous under specific combinations of conditions. The results suggest this method will be useful for monitoring burned area and burn severity in south-eastern USA vegetation types if the provided recommendations and limitations are considered.

1984–2010 trends in fire burn severity and area for the conterminous US

Burn severity products created by the Monitoring Trends in Burn Severity (MTBS) project were used to analyse historical trends in burn severity. Using a severity metric calculated by modelling the cumulative distribution of differenced Normalized Burn Ratio (dNBR) and Relativized dNBR (RdNBR) data, we examined burn area and burn severity of 4893 historical fires (1984–2010) distributed across the conterminous US (CONUS) and mapped by MTBS. Yearly mean burn severity values (weighted by area), maximum burn severity metric values, mean area of burn, maximum burn area and total burn area were evaluated within 27 US National Vegetation Classification macrogroups. Time series assessments of burned area and severity were performed using Mann–Kendall tests. Burned area and severity varied by vegetation classification, but most vegetation groups showed no detectable change during the 1984–2010 period. Of the 27 analysed vegetation groups, trend analysis revealed burned area increased in eight, and burn severity has increased in seven. This study suggests that burned area and severity, as measured by the severity metric based on dNBR or RdNBR, have not changed substantially for most vegetation groups evaluated within CONUS.

Timing Constraints on Remote Sensing of Wildland Fire Burned Area in the Southeastern US

Abstract: Remote sensing using Landsat Thematic Mapper (TM) satellite imagery is increasingly used for mapping wildland fire burned area and burn severity, owing to its frequency of collection, relatively high resolution, and availability free of charge. However, rapid response of vegetation following fire and frequent cloud cover pose challenges to this approach in the southeastern US. We assessed these timing constraints by using a series of Landsat TM images to determine how rapidly the remotely sensed burn scar signature fades following prescribed burns in wet flatwoods and depression swamp community types in the Apalachicola National Forest, Florida, USA during 2006. We used both the Normalized Burn Ratio (NBR) of reflectance bands sensitive to vegetation and exposed soil cover, as well as the change in NBR from before to after fire (dNBR), to estimate burned area. We also determined the average and maximum amount of time following fire required to obtain a cloud-free image for burns in each month of the year, as well as the predicted effect of this time lag on percent accuracy of burn scar estimates. Using both NBR and dNBR, the detectable area decreased linearly 9% per month on average over the first four months following fire. Our findings suggest that the NBR and dNBR methods for monitoring burned area in common southeastern US vegetation community types are limited to an average of 78–90% accuracy among months of the year, with individual burns having values as low as 38%, if restricted to use of Landsat 5 TM imagery. However, the majority of burns can still be mapped at accuracies similar to those

An Optimal Sample Data Usage Strategy to Minimize Overfitting and Underfitting Effects in Regression Tree Models Based on Remotely-Sensed Data

Regression tree models have been widely used for remote sensing-based ecosystem mapping. Improper use of the sample data (model training and testing data) may cause overfitting and underfitting effects in the model. The goal of this study is to develop an optimal sampling data usage strategy for any dataset and identify an appropriate number of rules in the regression tree model that will improve its accuracy and robustness. Landsat 8 data and Moderate-Resolution Imaging Spectroradiometer-scaled Normalized Difference Vegetation Index (NDVI) were used to develop regression tree models. A Python procedure was designed to generate random replications of model parameter options across a range of model development data sizes and rule number constraints. The mean absolute difference (MAD) between the predicted and actual NDVI (scaled NDVI, value from 0–200) and its variability across the different randomized replications were calculated to assess the accuracy and stability of the models. In our case study, a six-rule regression tree model developed from 80% of the sample data had the lowest MAD (MADtraining = 2.5 and MADtesting = 2.4), which was suggested as the optimal model. This study demonstrates how the training data and rule number selections impact model accuracy and provides important guidance for future remote-sensing-based ecosystem modeling.

Geospatial data mining for digital raster mapping

We performed an in-depth literature survey to identify the most popular data mining approaches that have been applied for raster mapping of ecological parameters through the use of Geographic Information Systems (GIS) and remotely sensed data. Popular data mining approaches included decision trees or “data mining” trees which consist of regression and classification trees, random forests, neural networks, and support vector machines. The advantages of each data mining approach as well as approaches to avoid overfitting are subsequently discussed. We also provide suggestions and examples for the mapping of problematic variables or classes, future or historical projections, and avoidance of model bias. Finally, we address the separate issues of parallel processing, error mapping, and incorporation of “no data” values into modeling processes. Given the improved availability of digital spatial products and remote sensing products, data mining approaches combined with parallel processing potentials should greatly improve the quality and extent of ecological datasets.