A model-based performance test for forest classifiers on remote-sensing imagery

Ambiguity between forest types on remote-sensing imagery is a major cause of errors found in accuracy assessments of forest inventory maps. This paper presents a methodology, based on forest plot inventory, ground measurements and simulated imagery, for systematically quantifying these ambiguities in the sense of the minimum distance (MD), maximum likelihood (ML), and frequency-based (FB) classifiers. The method is tested with multi-spectral IKONOS images acquired on areas containing six major communities (oak, pine, fir, primary and secondary high tropical forests, and avocado plantation) of the National Forest Inventory (NFI) map in Mexico. A structural record of the canopy and optical measurements (leaf area index and soil reflectance) were performed on one plot of each class. Intra-class signal variation was modelled using the Discrete Anisotropic Radiative Transfer (DART) simulator of remote-sensing images. Atmospheric conditions were inferred from ground measurements on reference surfaces and leaf optical properties of each forest type were derived from the IKONOS forest signal. Next, all forest types were simulated, using a common environmental configuration, in order to quantify similarity among all forest types, according to MD, ML and FB classifiers. Classes were considered ambiguous when their dissimilarity was smaller than intra-class signal variation.     

Estimating plot-level forest structural attributes using high spectral resolution ASTER satellite data in even-aged Eucalyptus plantations in southern KwaZulu-Natal,South Africa

This study assessed the suitability of both visible and shortwave infrared of ASTER reflectance bands and various vegetation indices for estimating forest structural attributes of Eucalyptus species. The study was conducted in even-aged monoculture plantations of E. grandis and E. nitens in the southern KwaZulu-Natal Midlands of South Africa. Empirical relationships between forest structural attributes, i.e. stems per hectare (SPHA), diameter at breast height (DBH), mean tree height (MTH), basal area and volume, and ASTER data were derived using correlation and canonical correlation analysis (CCA). The results indicated weak relationships between the studied forest structural attributes and ASTER data. In the younger plantation stands (4–6 years) the adjusted R ² values from CCA regression for SPHA, DBH, MTH, basal area and volume were 54.2, 63.5, 33.8, 25.4 and 30.3, respectively. The adjusted R ² values in the mature stands (7–9 years) were distinctly weaker with values of 50.7, 55.8, 25.1, 20.2 and 27.3 for SPHA, DBH, MTH, basal area and volume, respectively. The results imply that ASTER satellite data are not applicable to forest structural attribute estimation in commercially managed forest stands.     

Estimating above-ground biomass by fusion of LiDAR and multispectral data in subtropical woody plant communities in topographically complex terrain in North-eastern Australia

We investigated a strategy to improve predicting capacity of plot-scale above-ground biomass （AGB） by fusion of LiDAR and Land- sat5 TM derived biophysical variables for subtropical rainforest and eucalypts dominated forest in topographically complex landscapes in North-eastern Australia. Investigation was carried out in two study areas separately and in combination. From each plot of both study areas, LiDAR derived structural parameters of vegetation and reflectance of all Landsat bands, vegetation indices were employed. The regression analysis was carded out separately for LiDAR and Landsat derived variables indi- vidually and in combination. Strong relationships were found with LiDAR alone for eucalypts dominated forest and combined sites compared to the accuracy of AGB estimates by Landsat data. Fusing LiDAR with Landsat5 TM derived variables increased overall performance for the eucalypt forest and combined sites data by describing extra variation （3% for eucalypt forest and 2% combined sites） of field estimated plot-scale above-ground biomass. In contrast, separate LiDAR and imagery data, andfusion of LiDAR and Landsat data performed poorly across structurally complex closed canopy subtropical minforest. These findings reinforced that obtaining accurate estimates of above ground biomass using remotely sensed data is a function of the complexity of horizontal and vertical structural diversity of vegetation.     

Satellite-based stand-wise forest cover type mapping using a spatially adaptive classification approach

Due to high variation in forest communities, forest structure and the fragmentation of the forested area in Central Europe, satellite-based forest inventory methods have to meet particularly high-quality requirements. This study presents an innovative method to combine official forest inventory information at stand level with multidate satellite imagery using a spatially adaptive classification approach for producing wall-to-wall forest cover maps of important tree species and management classes across multiple ownership regions in a heterogeneous low mountain range in Germany. The classification approach was applied to a 5,200-km² area (about 2,080?km² of forest land, mostly mixed forests) located in the Eifel mountain range in Central Europe. In comparison with conventional classifiers, our results demonstrate a significant increase in classification accuracy in the order of 12%. The method was tested with ASTER images but holds the potential to be used for regular state forest inventories based on standard and novel earth observation data supplied for instance from the SPOT-5 and RapidEye sensors.     

Estimating models of vegetation fractional coverage based on remote sensing images at different radiometric correction levels

The images of post atmospheric correction reflectance (PAC), top of atmosphere reflectance (TOA), and digital number (DN) of a SPOT5 HRG remote sensing image of Nanjing, China were used to derive four vegetation indices (VIs), that is, normalized difference vegetation index (NDVI), transformed vegetation index (TVI), soil-adjusted vegetation index (SAVI), and modified soil-adjusted vegetation index (MSAVI). Based on these VIs and the vegetation fractional coverage (VFC) data obtained from field measurements, thirty-six VI-VFC relationship models were established. The results showed that cubic polynomial models based on NDVI and TVI from PAC were the best, followed by those based on SAVI and MSAVI from DN, with their accuracies being slightly higher than those of the former two models when VFC>0.8. The accuracies of these four models were higher in medium densely vegetated areas (VFC = 0.4–0.8) than in sparsely vegetated areas (VFC = 0–0.4). All the models could be used elsewhere via the introduction of a calibration model. In VI-VFC modeling, using VIs derived from different radiometric correction levels of remote sensing images could help explore and show valuable information from remote sensing data and thus improve the accuracy of VFC estimation.     

Vegetation inventory of a temperate biosphere reserve in China by image fusion of landsat TM and SPOT HRV

Vegetation cover types on Changbai Mountain, a natural biosphere reserve (2,000 km²) in northeast China, were derived by using multisensor satellite imagery fused with Landsat TM and SPOT HRV-XS. DEM data were used for improving classification accuracy. Cover types were classified into 20 groups. Bands 4 and 5 of Landsat TM image acquired on July 18, 1997, and band 1 of SPOT HRV-XS image acquired on Oct. 19, 1992, were fused to a false color image, and maximum likelihood supervised classification was performed. Data fusion showed high accuracy of identification, compared to individual images. The overall accuracy of classification of individual images by SPOT HRV-XS reached 56%, and TM 66%, while the fused data set provided accuracy of about 78%, which was raised to 81% after recoding by using DEM. There were five vegetation zones on the mountain, from the base to the peak: hardwood forest zone, mixed forest zone, conifer forest zone, birch forest zone, and tundra zone. Spruce-fir dominated conifer forest was the most prevalent (nearly 50%) vegetation type, followed by Korean pine and mixed forest (17%) and larch forest (5%). HRV image taken in leaf-off season is useful for discriminating forest from non-forest, and evergreen forest from hardwood forest, while the summer image (TM) provides detailed information on the difference in similar vegetation types, like hardwood forest with different compositions.     

Classifying tree species in a northern mixed forest using high-resolution IKONOS data

The high-spatial-resolution IKONOS satellite is now operating as a resource and disaster monitor, after a successful launch in September 1999. The ground resolution of the IKONOS panchromatic band is about 1m, the greatest of any satellite. The objectives of this study were to verify the extent to which high-resolution IKONOS data can be used to classify tree species. A field survey and image analysis study used IKONOS imagery to classify 21 species in mixed stands of deciduous and conifer species with the following results: (1) The panchromatic and multi-spectral bands 4, 3, and 2 were useful for classifying tree species owing to the great difference in the reflectance values between tree species. (2) Some groups, for which there were significant differences among species, were identified using Tukeys multiple comparison test; conifers and some broadleaved trees were identified correctly more often than other species. (3) A random selection of validation pixels showed that the overall classification accuracy was 62%. The classification accuracy of broadleaved trees was a little low, ranging from 40% to 63%, while that of conifers exceeded 70%. (4) The overall accuracy of the classification at the genus level improved by 4% more than the species level. The misclassification of broadleaved trees was due to the similar spectral characteristics of species in the same genus.     

Reliable, accurate and timely forest mapping for wildfire management using ASTER and Hyperion satellite imagery

The present study aims to explore the potential and effectiveness of new Earth Observation data for mapping the vegetation composition and structure and thus provide accurate forest maps to be used in fire propagation simulation models and fire risk assessment. Land cover classification of ASTER and Hyperion images is performed in a detailed nomenclature including different vegetation types and densities since the same vegetation type may give fires with different behaviour as a result of differences in fuel continuity.

The results suggest that both datasets can provide highly accurate maps with an overall accuracy of 85% for ASTER and 93% for Hyperion classification. Although Hyperion is superior to ASTER in terms of overall accuracy, the latter provided a higher thematic accuracy identifying one additional class compared to Hyperion. The evaluation of the classification results in terms of cost and technical characteristics suggest that both datasets are suitable for use in wildfire management tools, depending on the specific user needs, and they could also be used complementary if a combination of high thematic accuracy and locally high spatial accuracy is needed.     

Estimating vertical vegetation density through a SPOT5 imagery at multiple radiometric correction levels

There is growing concern about remote sensing of vertical vegetation density in rapidly expanding peri-urban interfaces.A widely used parameter for such density,i.e.,leaf area index (LAI),was measured in situ in Nanjing,China and then correlated with two vegetation indices (VI) derived from multiple radiometric correction levels of a SPOT5 imagery.The VIs were a normalized difference vegetation index (NDVI) and a ratio vegetation index (RVI),while the four radiometric correction levels were i) post atmospheric correction reflectance (PAC),ii) top of atmosphere reflectance (TOA),iii) satellite radiance (SR) and iv) digital number (DN).A total of 157 LAI-VI relationship models were established.The results showed that LAI is positively correlated with VI (r varies from 0.303 to 0.927,p < 0.001).The R 2 values of "pure" vegetation were generally higher than those of mixed vegetation.The average R 2 values of about 40 models based on DN data (0.688) were higher than that of the routinely used PAC (0.648).Independent variables of the optimal models for different vegetation quadrats included two vegetation indices at three radiometric correction levels,indicating the potential of vegetation indices at multiple radiometric correction levels in LAI inversion.The study demonstrates that taking heterogeneities of vegetation structures and uncertainties of radiometric corrections into account may help full mining of valuable information from remote sensing images,thus improving accuracies of LAI estimation.     

基于TM影像的荒漠林光谱特征及提取模型研究

以TM为信息源,对TM多光谱遥感影像的辐射校正、几何校正、研究区实地样地的标定、区划,以及识别模型的提取验证,分析荒漠林树种在TM影像上的机理,分析荒漠林树种光谱特征,建立基于光谱知识的荒漠林树种信息提取的识别模型。结果表明:TM图像的红柳与梭梭的光谱特征明显的差异主要集中在TM4波段上,红柳在TM4的亮度值明显高于梭梭,即红柳亮度值具有在TM4最大的特征;经过对模型的验证,模型判对率为70%。     

Estimation of LAI in the forested watershed using ASTER data based on Price’s model in summer and winter

Leaf area index (LAI) is an important parameter to identify the water balance in forested watershed as a biological factor influencing directly on the evapotranspiration in the forest area. The purpose of this study was to estimate the LAI in a small forested watershed in summer and winter by applying the Terra/Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) data to the LAI estimation method. In this study, the estimation was based on the absorption and scattering processes of the solar radiation in the vegetation canopy and the spectral reflectance characteristics of soil vegetation. First, we estimated LAI based on Price’s model by application of ASTER data on the forested watershed located in the Tenzan Mountains of Saga, Japan. To validate the results of LAI estimation, secondly, we compared them to the measured LAI obtained by a plant canopy analyzer (LAI-2000) on the observation area inside the target region. This study showed that the LAI estimation method was a feasible and accurate method as indicated by the high relationship (r = 0.97) between LAI derived from ASTER data and LAI measured by LAI-2000. This paper is the first report on LAI estimation using Terra/ASTER data based on Price’s model and field investigation. This LAI estimation method is a reliable and applicable method.     

Estimation of aboveground tree carbon stock using SPOT-HRG data (a case study: Darabkola forests)

Forests are among the most important carbon sinks on earth. However, their complex structure and vast areas preclude accurate estimation of forest carbon stocks. Data sets from forest monitoring using advanced satellite imagery are now used in international policy agreements. Data sets enable tracking of emissions of CO₂ into the atmosphere caused by deforestation and other types of land-use changes. The aim of this study is to determine the capability of SPOT-HRG Satellite data to estimate aboveground carbon stock in a district of Darabkola research and training forest, Iran. Preprocessing to eliminate or reduce geometric error and atmospheric error were performed on the images. Using cluster sampling, 165 sample plots were taken. Of 165 plots, 81 were in natural habitats, and 84 were in forest plantations. Following the collection of ground data, biomass and carbon stocks were quantified for the sample plots on a per hectare basis. Nonparametric regression models such as support vector regression were used for modeling purposes with different kernels including linear, sigmoid, polynomial, and radial basis function. The results showed that a third-degree polynomial was the best model for the entire studied areas having an root mean square error, bias and accuracy, respectively, of 38.41, 5.31, and 62.2; 42.77, 16.58, and 57.3% for the best polynomial for natural forest; and 44.71, 2.31, and 64.3% for afforestation. Overall, these results indicate that SPOT-HRG satellite data and support vector machines are useful for estimating aboveground carbon stock.     

Characterizing rangeland vegetation using Landsat and 1-mm VLSA data in central Wyoming (USA)

As an alternative to ground-cover data collection by conventional and expensive sampling techniques, we compared measurements obtained from very large scale aerial (VLSA) imagery for calibrating moderate resolution Landsat data. Using a grid-based sampling scheme, 162 VLSA images were acquired at 100 m above ground level. The percent vegetation cover in each photo was derived using SamplePoint (a manual inventory method) and VegMeasure (a reflectance based, automated method). Approximately two-thirds of the VLSA images were used for calibrating Landsat data while the remainder was used for validation. Regression models with Landsat bands accounted for 55% of the VegMeasure-based measurements of vegetation, whereas models that included both Landsat bands and elevation data accounted for 67%. The relationship between the Landsat bands and the percent vegetation cover measured by SamplePoint was lower (R ² = 20%), highlighting the differences between the inventory and reflectance based protocols. Results from the model validation indicated that the model’s predictive power was lower when the vegetation cover was either <20% or >55%. Additional work is needed in these ecosystems to improve the calibration techniques for sites with low and high vegetation cover; however, these results demonstrate the VLSA imagery could be used for calibrating Landsat data and deriving rangeland vegetation cover. By adopting such methodologies the US Federal land management agencies can increase the efficiency of the monitoring programs in Wyoming and in other western states of the US. Mention of trade names is for information only and does not imply endorsement by USDA over comparable products or services.     

Using multitemporal satellite imagery to characterize forest wildlife habitat: The case of ruffed grouse

Like many similar forest species, ruffed grouse (Bonasa umbellus; hereafter grouse) populations in the central and southern Appalachians (CSA) are strongly affected by forest composition at the landscape scale. Because these populations are in decline, managers require accurate forest maps to understand how stand level characteristics affect the survival and reproductive potentials of individual birds to design management strategies that improve grouse abundance. However, traditional mapping techniques are often labor-intensive and cost-prohibitive. We used a normalized difference vegetation index (NDVI) from each of 8 Landsat images and the digital elevation model (DEM)-derived variables of elevation and aspect in discriminant analyses to classify 7 study areas to 3 overstory classes (evergreen, hardwoods, and oak) and distinguish evergreen and deciduous understories in the CSA, 2000–2002. Overall accuracy was 82.08%, varying from 83.59% for oak to 79.79% for hardwoods overstories. Periods with large phenological differences among classes, particularly early and late spring, were most useful for discriminating overstory vegetation types. Alternatively, winter NDVI in combination with elevation was critical for differentiating evergreen and deciduous understories. Multitemporal image sets used in concert with DEMs provided a cost-effective alternative to hyperspectral sensors for improving wildlife habitat classification accuracy with Landsat imagery. This allowed for enhanced understanding of grouse-habitat relationships and habitat affects on grouse populations that allowed for improved management. With the incorporation of simple adjustments for local forest plant species phenology into the model, it may be used to better classify wildlife habitat of similar species in areas with comparable forest communities and topography. Multitemporal images can also be used to differentiate grassland communities, monitor wetlands, and serve as baseline data for detecting changes in land use over longer temporal scales, making their use in forest wildlife habitat studies cost-justifiable.     

Forest stand age classification using time series of photogrammetrically derived digital surface models

In this research, we developed and tested a remote sensing-based approach for stand age estimation. The approach is based on changes in the forest canopy height measured from a time series of photo-based digital surface models that were normalized to canopy height models using an airborne laser scanning derived digital terrain model (DTM). Representing the Karelian countryside, Finland, CHMs from 1944, 1959, 1965, 1977, 1983, 1991, 2003, and 2012 were generated and allow for characterization of forest structure over a 68-year period. To validate our method, we measured stand age from 90 plots (1256?m²) in 2014, whereby producer's accuracy ranged from 25.0% to 100.0% and user's accuracy from 16.7% to 100.0%. The wide range of accuracy found is largely attributable to the quality and characteristics of archival images and intrastand variation in stand age. The lowest classification accuracies were obtained for the images representing the earliest dates. For forest managers and agencies that have access to long-term photo archives and a detailed DTM, the estimation of stand age can be performed, improving the quality and completeness of forest inventory databases.     

基于Prospect,Liberty和Geosail模型的森林叶面积指数的反演

针对传统的统计模型方法反演叶面积指数(LAI)具有不稳定、区域不统一性的缺点,本研究从物理机制角度出发,以Prospect,Liberty和Geosail模型为基础,建立查找表从TM影像上反演LAI,并与TRAC实测的LAI比较。结果表明:基于机制模型与查找表的方法反演的LAI与实测的LAI有较好的一致性,实测精度达到83.7%。     

不同地面控制点采集方法对SPOT5图像几何精校正精度的影响

近年来,SPOT5卫星遥感数据在林业上的应用越来越广泛,由于林区缺乏大比例尺地形图,图像几何精校正存在一些困难,因此有必要探讨一种科学可靠且适宜于大面积应用的方法。通过1：10000地形图和航空正射图像采集地面控制点（GCPs）,结合林区1：50000DEM数据和SPOT5物理模型对图像进行2维平面校正和3维正射校正的方法试验,结果表明：航空图像采集GCPs进行平面和正射校正的精度都高于地形图采集GCPs的平面和正射校正精度;从地形图采集GCPs正射校正精度高于平面校正精度。从航空图像采集GCPs正射校正精度低于平面校正精度。由于精度高、获取方便且价格可以接受,航空图像采集GCPs进行SPOT5图像几何精校正是当前大面积应用中较适宜的方法。     

基于时序Landsat数据的浙江省竹林信息提取及时空演变

【目的】提取浙江省不同时期竹林分布信息,分析其时空演变规律,揭示竹林面积变化与土地利用格局之间的关系,为国家及至全球尺度长时间序列的竹林时空动态研究提供参考。【方法】以浙江省为研究区,基于2000、2004和2008年Landsat5 TM及2014年Landast8 OLI时间序列影像数据,首先,对不同时期的Landsat数据进行大气校正和几何校正,采用最大似然法提取土地利用和竹林时空分布信息;然后,利用变化幅度和动态度2个指标分析4个时期、3个时间段的竹林时空演变规律;最后,建立全省土地利用时空转移矩阵,揭示竹林时空动态与土地利用格局之间的关系。【结果】1)基于时序Landsat数据提取的浙江省竹林信息精度较高,分类精度达75%以上,使用者精度达91%以上,且分类统计面积与实际清查面积高度吻合,面积提取精度达96%以上; 2) 2000—2014年浙江省竹林面积变化幅度和年均变化率分别为16.55%和1.18%,在时空上呈逐渐增加趋势; 3)浙江省竹林面积由2000年占全省面积的7.33%增长到2014年的8.56%,其中针叶林、阔叶林和农田3种土地利用类型变化对竹林面积增加的贡献最大,贡献率分别为28.62%、37.23%和16.15%。【结论】基于Landsat时间序列数据能够高精度监测浙江省竹林资源动态变化,针叶林、阔叶林和农田等土地利用类型减少对竹林面积时空演变具有显著影响。     

Forecasting tree mortality using change metrics derived from MODIS satellite data

Insect-induced tree mortality can cause substantial timber and carbon losses in many regions of the world. There is a critical need to forecast tree mortality to guide forest management decisions. Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery provides inexpensive and frequent coverage over large areas, facilitating forest health monitoring. This study examined time series of MODIS satellite images to forecast tree mortality for a Pinus radiata plantation in southern New South Wales, Australia. Dead tree density derived from ADS40 aerial imagery was used to evaluate the performance of change metrics derived from time series of MODIS-based vegetation indices. Continuous subset selection by LASSO regression and model assessment using a variant of the bootstrap were used to select the best performing change metrics out of a large amount of predictor variables to account for over-fitting. The results suggest that 250 m 16-daily MODIS images are effective for forecasting tree mortality. Seasonal change metrics derived from the Normalized Difference Vegetation Index (NDVI) outperformed the Enhanced Vegetation Index (EVI) and the Normalized Difference Infrared Index (NDII). Temporal analysis illustrated that optimal forecasting power was obtained using change metrics based on three years of satellite data for this population. The forecast could be used to optimise the scheduling of detailed forest health surveys and silvicultural operations which currently are planned based on stratified, annual assessments. This coarse-scale, spatio-temporal analysis represents a potentially cost-effective early warning approach to forecasting tree mortality in pine plantations by identifying compartments that require more detailed investigation.     

Spatial partitioning of biomass and diversity in a lowland Bolivian forest: Linking field and remote sensing measurements

Large-scale inventories of forest biomass and structure are necessary for both understanding carbon dynamics and conserving biodiversity. High-resolution satellite imagery is starting to enable structural analysis of tropical forests over large areas, but we lack an understanding of how tropical forest biomass links to remote sensing. We quantified the spatial distribution of biomass and tree species diversity over 4 ha in a Bolivian lowland moist tropical forest, and then linked our field measurements to high-resolution Quickbird satellite imagery. Our field measurements showed that emergent and canopy dominant trees, being those directly visible from nadir remote sensors, comprised the highest diversity of tree species, represented 86% of all tree species found in our study plots, and contained the majority of forest biomass. Emergent trees obscured 1–15 trees with trunk diameters (at 1.3 m, diameter at breast height (DBH)) ≥20 cm, thus hiding 30–50% of forest biomass from nadir viewing. Allometric equations were developed to link remotely visible crown features to stand parameters, showing that the maximum tree crown length explains 50–70% of the individual tree biomass. We then developed correction equations to derive aboveground forest biomass, basal area, and tree density from tree crowns visible to nadir satellites. We applied an automated tree crown delineation procedure to a high-resolution panchromatic Quickbird image of our study area, which showed promise for identification of forest biomass at community scales, but which also highlighted the difficulties of remotely sensing forest structure at the individual tree level.