Progress 06/01/14 to 01/31/16
Outputs
Target Audience:Advancing satellite remote sensing technologies such as Synthetic Aperture Radar (SAR) have been shown in many research applications to be capable of addressing information gaps unable to be filled by the use of optical sensors, such as the well-known Landsat constellation. SAR sensors are capable of penetrating canopies, observing in all weather conditions, and have sensitivity to key forest biophysical parameters. The sensitivity of SAR to structural information (biomass, density, vertical layering, spacing) makes SAR advantageous, especially when cloud cover is problematic. Many non-experts are becoming aware of Landsat-based forest monitoring products since the transition to open archives and the launch of tools like Google Earth Engine. For example, Matt Hanson's forest maps and platforms such as Global Forest Watch are changing the way decisions get made. With the release of PALSAR-1 L-band data and open policy of Sentinel-1 C-band data, now collecting operationally and free, the opportunity to develop operational SAR forest monitoring is enormous, however challenges remain. The wide adaptation of SAR data and accompanying techniques for disturbance mapping has not occurred primarily due to: complexity of data formats and challenges of data handling lack of operational algorithms for generating forest attribute metrics lack of high resolution Digital Elevation Models (DEM) required for processing cost, lack of coverage, and baseline observations scenarios not wall-to-wall collection challenges of data fusion and multitemporal analyses In a way these technical barriers have created an innovation gap and opportunity. Keep in mind that Landsat, which tends to be more intuitive since it covers the visible portion of the spectrum, has been around for nearly 50 years. While SAR technology has a long history its use has remained elusive. However, many of the early barriers to SAR development are falling. Today, DEMs are readily available anywhere and constantly improving. Space agencies are shifting toward open data policies and distribute free L- and C- band satellite data. Observation strategies are now wall-to-wall with improving resolutions and have thus created "BigData" opportunities. Advancing computational techniques and cloud computing are increasing access and transferring information to the masses. Many "research studies" have executed projects using SAR and space agencies are now beginning to democratize science. The next step is to advance the tools and science for operational algorithms to generate forest characteristic products and mapping services by taking advantage of new SAR missions. This will greatly enhance forest monitoring and improve decision making. Over the next decade a suite of even more SAR sensors (e.g., NISAR, BIOMASS) with more operational power are planned for launch. This SBIR Phase II addresses this exact opportunity and will provide the foundation for success. United States Department of Agriculture (USDA) United States Forest Service (USFS)- Forest lands comprise 304million hectares of the total land area in the US with 44% controlled by Federal, State, and local governments. Characterizing forest structure, quantifying forest disturbance, and monitoring forest utilization on these lands is a primary and continuous need by multiple end user communities, including the USDA USFS (see letters). Both natural and anthropogenic stressors are influencing forest disturbance processes and patterns. Climate change, extreme weather, renewable energy demands, timber, land conversion, and invasive species are several key actors influencing forest function and services. These can range in frequency, intensity, and magnitude. Timely and accurate assessment of disturbance events and condition are critical to respond effectively and appropriately at local, regional, and national scales. In addition, assessing the severity and understanding the scales of disturbance are critical for broad-scale management, policy, and economic decision-making. While optical-based mapping techniques have made significant contributions toward assessment, they are limited by cloud coverage and only provided limited information on stand structure. The wide adaptation of operational SAR derived metrics into USFS programs and disturbance protocol has remained elusive due to the technical barriers. In this SBIR we intend to bridge this gap and provide enhanced products and services (metrics, maps, techniques, and expertise) that are easy to apply and integrate into decision support tools. Monitoring, Reporting, and Verification (MRV) community- The rapid growth of carbon science and characterization of ecosystem services has created a huge market for tools, products, and services. International monitoring frameworks, such as Reducing Emissions from Deforestation and Degradation (REDD), have created the need for robust Monitoring, Reporting, and Verification (MRV) platforms that can serve programs and initiatives aimed at carbon cycle science, conservation, and conventions. Maps of deforestation and degradation at multiple scales have taken on increased importance with the recognition of the value of forests and the rates of forest loss. International programs, like REDD, are being constructed to help preserve the world's forests, improve livelihoods, and promote sustainable practices. Under REDD, emissions and removals of carbon due to changes in forests are estimated from two components: activity data (e.g. deforestation, degradation) and forest carbon stock data (e.g. carbon in biomass, soil c). As systems are built in which some people pay others to slow the rate of carbon emissions from forest loss, the need for accurate, spatially explicit, and timely information on forests and forest change becomes critical. Satellite observations have been used to map forests since the 1970s. While optical reflectance data are the most abundant and common type of observations available, radar and Lidar data can be used to effectively map forest characteristics and changes. Optical reflectance observations passively measure the sun's radiation off of the land surface. Trees, forested areas, and changes in these areas are identified based on the amount and change in reflectance. SAR and Lidar are active systems that send out energy to the Earth's surface and measure the subsequent energy return. These active systems often provide more information regarding forest structure and, therefore, are useful for biomass and carbon content. Additionally, SAR systems are much less sensitive to cloud and haze cover than are optical systems, making them useful for forest monitoring in cloudy areas. However, many MRV/REDD platforms do not have the capacity to integrate SAR forest metrics. In this SBIR, we have strategically aligned the goals to serve this community (see letters). Our small business (AGS) has several high profile example MRV or "Decision Support Tools" (DST) currently operating that blend satellite maps, geospatial information, graphing and tabular data, mobile Apps, and end user options within web-GIS. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?Training and workshops to partners and end users. How have the results been disseminated to communities of interest?Yes What do you plan to do during the next reporting period to accomplish the goals?
Nothing Reported
Impacts
What was accomplished under these goals?
We determined RFDI was not optimal for our "low impact" case study disturbance (minor infestation) in the northeast. However, we note the Asian Longhorn Beetle "field data" from partners was not ideal given the lack of positional accuracy and quantitative metrics. Therefore, we also tried backscatter thresholding and examining multitemporal changes in sigma nought while comparing to optical indices such as Landsat NDVI. During Phase II we will co-collect this data with partners. We note that scale is a key element throughout this work and therefore emphasize the need to consider multiscale SAR data linked to the scale (or minimum mapping unit) of the field data AND issue to truly understand the strengths and limitations of detecting varying disturbance levels in an automated approach. The Phase II plan will address these scales directly with quantifiable metrics and field data. We also will support USFS programs with these metrics during the Phase II. We know we can automate RFDI for "high impact" case studies and scaling is not an issue (ie, CONUS product feasible) and we will continue to work with ForWarn. During Phase I we also tested JAXA's "Forest Non-Forest" (FNF) Mosaics generated from a decision tree approach. We created before and after products for disturbances. The FNF maps did poorly in mapping the disturbance demonstration case studies. A conclusion during Phase I was that FNF maps need to be locally tuned using sigma nought values as the "global" calibration was not sensitive to changes that were on the order of ~4db resulting from the tornado disturbance. RFDI was not useful for predicting AGB, or CCC in our case studies. During Phase II we will focus on locally tuned indices, multitemporal change vectors analyses, and add Sentinel-1 C-band data in addition to PALSAR-2 for a more thorough and robust toolset. Summarizing a case study, during Phase I, we evaluated empirical and semi-analytical models to generate SAR derived metrics of AGB, CCC, basal area, and disturbance. To evaluate performance we used a suite of statistics (R2, RMSE, Q-Q plots, residuals, Cook's distance, AIC, etc...) and expert judgement while considering cost and approach. For CCC and AGB we relied on semi-empirical regression using our compiled field sites across the region. We considered the explanatory power of the polarizations, ease of model execution, and benefits of more straightforward models vs models with higher accuracy but more complexity to pick an optimal approach for mapping attributes at the demonstration sites. Using the multicriteria approach the optimal models underwent n-folds cross validation to provide out of sample evaluation statistics (i.e., goodness of fit, RMSE) for predictions. During Phase I we achieved target accuracy goals for our AGB, basal area, and CCC metrics with R2 or 0.74, 0.79, and 0.75, respectively. Below we highlight AGB and CCC scatterplots of the cross validation (fig 6) where we mapped at 12.5m spatial resolution. We then extrapolated the models to the demonstration sites and region to help with assessment. During Phase I we also tested processing / modeling approaches including Water-Cloud and Random Volume over Ground. During Phase II we will enhance the number of field sites and integrate FIA data to expand regions with tuned estimates. Further, with this increase inn we can scale using more robust modeling approaches (i.e., physical retrieval models) as we generalize and integrate Sentinel-1 and Landsat 8 fusion. During Phase I our assessment concluded that in Carroll County more than 2500 hectares had modest (>50%) canopy damage and 1410 hectares had severe damage with a mean AGB loss of ~114 Mg / hectares in that severe damaged path. Total AGB loss in Carroll County was approximately 298,250Mg from the tornado (fig 7). From a cost perspective this would be a fraction of airborne Lidar (0.01 / hectare) considering the scalability. Phase 1 InSAR Forest Stand Height (FSH) During Phase I we tested repeat-pass InSAR techniques to map FSH at sites in the WMNH, NH sea coast, and across the entire northeast. Repeat-pass InSAR correlation measurements, assuming that the ground scattering effects can be ignored for cross-polarized SAR returns and also the interferometric vertical wavenumber, , is small (e.g., <0.05 rad/m), is the temporal change effects that dominate the repeat-pass InSAR correlation. The temporal change effects here refer to both the moisture change and random motion of the tree components during the repeat period of the satellite. As long as vegetation height dependence of the random motion effect, it is possible to utilize this dependence to invert forest height. In particular, the cross polarized repeat-pass InSAR correlation magnitude, , is connected to the physical forest height, , by the following relationship; where (from 0 to 1; unitless) is the model parameter accounting for moisture change and (>0; in the unit of meters) relates to random motion. We scaled up a forest stand height map (left) using repeat-pass InSAR using WMNF lidar. The WMNF scatterplot highlights InSAR inverted FSH and GRANIT Lidar-derived canopy height at 200m x 200m scale. We used more than 200 PALSAR-1 pairs to refine down to the optimal 36 InSAR scenes identified by coherence values for generating the map. We did notice weather driven moisture issues that we plan to address in Phase II. Python code was built to automate this process which is theoretically scalable to any sized region, such as CONUS. Stand scale or ~5 hectare minimum mapping unit is more appropriate than pixel level in this case.
Publications
|