Heather Reese

Countrywide Estimates of Forest Variables Using Satellite Data and Field Data from the National Forest Inventory

Abstract About 55% of the land area in Sweden is covered by forest. Presently, there is no uniform and geographically explicit description of the forest cover for Sweden. Independent data sources already existing are satellite data from Landsat TM and SPOT HRV, map masks and forest inventory plots. Together, they provide the possibility of computing estimates of forest variables such as stem volume and stand age. The accuracy for these estimations will be low on the pixel level, but higher on a stand level. This type of raster-based forest data is useful for authorities when planning how timber resources should be utilized or for monitoring purposes. The production line that was developed to create a countrywide database of forest variable estimates in a cost-efficient way is described. Accuracy of forest variable estimates for an area in southwestern Sweden was assessed at stand level. Results showed 33% overall root mean square error for the estimates of total wood volume, and 23% for the age estimates.

Applications using estimates of forest parameters derived from satellite and forest inventory data

From the combination of optical satellite data, digital map data, and forest inventory plot data, continuous estimates have been made for several forest parameters (wood volume, age and biomass). Five different project areas within Sweden are presented which have utilized these estimates for a range of applications. The method for estimating the forest parameters was a ”k-Nearest Neighbor” algorithm, which used a weighted mean value of k spectrally similar reference plots. Reference data were obtained from the Swedish National Forest Inventory. The output was continuous estimates at the pixel level for each of the variables estimated. Validation results show that accuracy of the estimates for all parameters was low at the pixel level (e.g., for total wood volume RMSE ranged from 58-80%), with a tendency toward the mean, and an underestimation of higher values while overestimating lower values. However, when the accuracy of the estimates is assessed over larger areas, the errors are lower, with best results being 10% RMSE over a 100 ha aggregation, and 17% RMSE over a 19 ha aggregation. Applications presented in this paper include moose and bird habitat studies, county level planning activities, use as input information to prognostic programs, and computation of statistics on timber volume within drainage basins and smaller land holdings. This paper provides a background on the kNN method and gives examples of how end users are currently applying satellite-produced estimation data such as these.

Mapping Tree Canopy Cover and Aboveground Biomass in Sudano-Sahelian Woodlands Using Landsat 8 and Random Forest

Accurate and timely maps of tree cover attributes are important tools for environmental research and natural resource management. We evaluate the utility of Landsat 8 for mapping tree canopy cover (TCC) and aboveground biomass (AGB) in a woodland landscape in Burkina Faso. Field data and WorldView-2 imagery were used to assemble the reference dataset. Spectral, texture, and phenology predictor variables were extracted from Landsat 8 imagery and used as input to Random Forest (RF) models. RF models based on multi-temporal and single date imagery were compared to determine the influence of phenology predictor variables. The effect of reducing the number of predictor variables on the RF predictions was also investigated. The model error was assessed using 10-fold cross validation. The most accurate models were created using multi-temporal imagery and variable selection, for both TCC (five predictor variables) and AGB (four predictor variables). The coefficient of determination of predicted versus observed values was 0.77 for TCC (RMSE = 8.9%) and 0.57 for AGB (RMSE = 17.6 tons∙ha−1). This mapping approach is based on freely available Landsat 8 data and relatively simple analytical methods, and is therefore applicable in woodland areas where sufficient reference data are available.

Combining airborne laser scanning data and optical satellite data for classification of alpine vegetation

Abstract Climate change and outdated vegetation maps are among the reasons for renewed interest in mapping sensitive alpine and subalpine vegetation. Satellite data combined with elevation derivatives have been shown to be useful for mapping alpine vegetation, however, there is room for improvement. The inclusion of airborne laser scanning data metrics has not been widely investigated for alpine vegetation. This study has combined SPOT 5 satellite data, elevation derivatives, and laser data metrics for a 25xa0kmxa0×xa031xa0km study area in Abisko, Sweden. Nine detailed vegetation classes defined by height, density and species composition in addition to snow/ice, water, and bare rock were classified using a supervised Random Forest classifier. Several of the classes consisted of shrub and grass species with a maximum height of 0.4xa0m or less. Laser data metrics were calculated from the nDSM based on a 10xa0mxa0×xa010xa0m grid, and after variable selection, the metrics used in the classification were the 95th and 99th height percentiles, a vertical canopy density metric, the mean and standard deviation of height, a vegetation ratio based on the raw laser data point cloud with a variable height threshold (from 0.1 to 1.0xa0m with 0.1xa0m intervals), and standard deviation of these vegetation ratios. The satellite data used in classification was all SPOT bands plus NDVI and NDII, while the elevation derivatives consisted of elevation, slope and the Saga Wetness Index. Overall accuracy when using the combination of laser data metrics, elevation derivatives and SPOT 5 data increased by 6% as compared to classification of SPOT and elevation derivatives only, and increased by 14.2% compared to SPOT 5 data alone. The classes which benefitted most from inclusion of laser data metrics were mountain birch and alpine willow. The producers accuracy for willow increased from 18% (SPOT alone) to 41% (SPOTxa0+xa0elevation derivatives) and then to 55% (SPOTxa0+xa0elevation derivativesxa0+xa0laser data) when laser data were included, with the 95th height percentile and Saga Wetness Index contributing most to willows improved classification. Addition of laser data metrics did not increase the classification accuracy of spectrally similar dry heath (

Tree Crown Mapping in Managed Woodlands (Parklands) of Semi-Arid West Africa Using WorldView-2 Imagery and Geographic Object Based Image Analysis

Detailed information on tree cover structure is critical for research and monitoring programs targeting African woodlands, including agroforestry parklands. High spatial resolution satellite imagery represents a potentially effective alternative to field-based surveys, but requires the development of accurate methods to automate information extraction. This study presents a method for tree crown mapping based on Geographic Object Based Image Analysis (GEOBIA) that use spectral and geometric information to detect and delineate individual tree crowns and crown clusters. The method was implemented on a WorldView-2 image acquired over the parklands of Saponé, Burkina Faso, and rigorously evaluated against field reference data. The overall detection rate was 85.4% for individual tree crowns and crown clusters, with lower accuracies in areas with high tree density and dense understory vegetation. The overall delineation error (expressed as the difference between area of delineated object and crown area measured in the field) was 45.6% for individual tree crowns and 61.5% for crown clusters. Delineation accuracies were higher for medium (35–100 m2) and large (≥100 m2) trees compared to small (<35 m2) trees. The results indicate potential of GEOBIA and WorldView-2 imagery for tree crown mapping in parkland landscapes and similar woodland areas.

Comparison of Manual Mapping and Automated Object-Based Image Analysis of Non-Submerged Aquatic Vegetation from Very-High-Resolution UAS Images

Aquatic vegetation has important ecological and regulatory functions and should be monitored in order to detect ecosystem changes. Field data collection is often costly and time-consuming; remote sensing with unmanned aircraft systems (UASs) provides aerial images with sub-decimetre resolution and offers a potential data source for vegetation mapping. In a manual mapping approach, UAS true-colour images with 5-cm-resolution pixels allowed for the identification of non-submerged aquatic vegetation at the species level. However, manual mapping is labour-intensive, and while automated classification methods are available, they have rarely been evaluated for aquatic vegetation, particularly at the scale of individual vegetation stands. We evaluated classification accuracy and time-efficiency for mapping non-submerged aquatic vegetation at three levels of detail at five test sites (100 m × 100 m) differing in vegetation complexity. We used object-based image analysis and tested two classification methods (threshold classification and Random Forest) using eCognition®. The automated classification results were compared to results from manual mapping. Using threshold classification, overall accuracy at the five test sites ranged from 93% to 99% for the water-versus-vegetation level and from 62% to 90% for the growth-form level. Using Random Forest classification, overall accuracy ranged from 56% to 94% for the growth-form level and from 52% to 75% for the dominant-taxon level. Overall classification accuracy decreased with increasing vegetation complexity. In test sites with more complex vegetation, automated classification was more time-efficient than manual mapping. This study demonstrated that automated classification of non-submerged aquatic vegetation from true-colour UAS images was feasible, indicating good potential for operative mapping of aquatic vegetation. When choosing the preferred mapping method (manual versus automated) the desired level of thematic detail and the required accuracy for the mapping task needs to be considered.

Mapping ground lichens using forest inventory and optical satellite data

Lichen is a major forage resource for reindeer and may constitute up to 80% of a reindeers winter diet. The reindeer grazing area in Sweden covers almost half of the country, with reindeer using mountainous areas in the summer and forested areas in the winter. Knowledge about the spatial distribution of ground lichens is important for both practical and decision-making purposes. Since the early 1980s, remote sensing research of lichen cover in northern environments has focused on reindeer grazing issues. The objective of this study was to use lichen information collected in the Swedish National Forest Inventory (NFI) as training data to classify optical satellite images into ground lichen cover classes. The study site was located within the reindeer husbandry area in northern Sweden and consisted of the common area between two contiguous Satellite Pour lObservation de la Terre (SPOT)-5 scenes and one Landsat-7 Enhanced Thematic Mapper Plus (ETM+) scene. Three classification methods were tested: Mahalanobis distance, maximum likelihood and spectral mixture analysis. Post-classification calibration was applied using a membership probability threshold in order to match the NFI-measured proportions of lichen coverage classes. The classification results were assessed using an independently collected field dataset (229 validation areas). The results demonstrated high classification accuracy of SPOT imagery for the classification of lichen-abundant and lichen-poor areas when using the Mahalanobis distance classifier (overall accuracy 84.3%, kappau2009=u20090.68). The highest classification accuracy for Landsat was achieved using a maximum likelihood classification (overall accuracy 76.8%, kappau2009=u20090.53). These results provided an initial indication of the utility of NFI data as training data in the process of mapping lichen classes over large areas.

Assessing the potential of multi-seasonal WorldView-2 imagery for mapping West African agroforestry tree species

High resolution satellite systems enable efficient and detailed mapping of tree cover, with high potential to support both natural resource monitoring and ecological research. This study investigates the capability of multi-seasonal WorldView-2 imagery to map five dominant tree species at the individual tree crown level in a parkland landscape in central Burkina Faso. The Random Forest algorithm is used for object based tree species classification and for assessing the relative importance of WorldView-2 predictors. The classification accuracies from using wet season, dry season and multi-seasonal datasets are compared to gain insights about the optimal timing for image acquisition. The multi-seasonal dataset produced the most accurate classifications, with an overall accuracy (OA) of 83.4%. For classifications based on single date imagery, the dry season (OA = 78.4%) proved to be more suitable than the wet season (OA = 68.1%). The predictors that contributed most to the classification success were based on the red edge band and visible wavelengths, in particular green and yellow. It was therefore concluded that WorldView- 2, with its unique band configuration, represents a suitable data source for tree species mapping in West African parklands. These results are particularly promising when considering the recently launched WorldView-3, which provides data both at higher spatial and spectral resolution, including shortwave infrared bands.

Combining point clouds from image matching with SPOT 5 multispectral data for mountain vegetation classification

There is a need to replace outdated vegetation maps over Sweden’s mountain region; the ability and accuracy of creating such maps with automated methods and remotely sensed data has been a topic of recent research. While spectral information is a key data input for mapping mountain vegetation, the addition of three-dimensional (3D) data has also proven useful in classification. Point clouds from photogrammetric image matching (IM) or from airborne laser scanning (ALS) are potential 3D data sources. In this study, vegetation height and density metrics from IM and ALS data were classified both alone and in combination with SPOT 5 (Système Probatoire d’Observation de la Terre) satellite data and elevation data (elevation, slope, and a wetness index). A Random Forest classification was used to map alpine and subalpine vegetation over Abisko, Sweden. The most notable result in this study was higher producer’s accuracy of the mountain birch classification when using IM metrics alone (98%) as compared to ALS data alone (89%). Classification of IM, SPOT, and elevation data combined gave the same overall accuracy (83%) as when using ALS, SPOT, and elevation data combined (also 83%). While most of the alpine vegetation classes were poorly classified using either the IM or ALS metrics alone, the IM point cloud appeared to contain more information for lower-growing (<2 m) vegetation than the ALS point cloud.

Forest mapping using 3D data from SPOT-5 HRS and Z/I DMC

The nation-wide Airborne Laser Scanning (ALS) currently performed by the Swedish National Land Survey will provide a new and accurate Digital Elevation Model (DEM). These data will enable new and cost-efficient assessments of vegetation height using Canopy Height Models (CHMs) derived as the difference between a Digital Surface Model (DSM) and the DEM. In this context, the High Resolution Stereoscopic (HRS) sensor onboard SPOT-5 and the airborne Z/I Digital Mapping Camera (DMC) used for operational aerial photography by the Swedish National Land Survey are of main interest. Previous research has shown that reliable tree height data are a powerful source of information for forest management planning. This study investigated the possibilities to map forest variables using CHMs derived from either the SPOT-5 HRS or Z/I DMC sensor together with ALS DEM data, in combination with spectral data from the SPOT-5 High Resolution Geometric (HRG) sensor. The results when using the Z/I DMC CHM in combination with SPOT-5 HRG data showed Root Mean Square Errors for standwise prediction of mean tree height, stem diameter, and stem volume of 7.3%, 9.0%, and 19%, respectively. The SPOT-5 HRS CHM in combination with SPOT-5 HRG data improved the SPOT HRG based estimates from 13% to 10%, 15% to 13%, and 31% to 23%, for tree height, stem diameter, and stem volume, respectively. Adding CHM data to a SPOT-5 HRG based prediction model improved the mapping accuracy between 13% to 44%. In conclusion, the obtained accuracies may be sufficient for operational forest management planning.