A two-step nearest neighbors algorithm using satellite imagery for predicting forest structure within species composition classes

Nearest neighbors techniques have been shown to be useful for predicting multiple forest attributes from forest inventory and Landsat satellite image data. However, in regions lacking good digital land cover information, nearest neighbors selected to predict continuous variables such as tree volume must be selected without regard to relevant categorical variables such as forest/non-forest. The result is that non-zero volume predictions may be obtained for pixels predicted to be non-forest, and volume predictions for pixels predicted to be forest may be erroneously small due to non-forest nearest neighbors. For users who wish to circumvent this discrepancy, a two-step algorithm is proposed in which the class of a relevant categorical variable such as land cover is predicted in the first step, and continuous variables such as volume are predicted in the second step subject to the constraint that all nearest neighbors must come from the predicted class of the categorical variable. Nearest neighbors, multinomial logistic regression, and discriminant analysis techniques were investigated for use in the first step. The results were generally similar for the three techniques, although the multinomial logistic regression technique was slightly superior. The k-Nearest Neighbors technique was used in the second step because many continuous forest inventory variables do not satisfy the distributional assumptions necessary for parametric multivariate techniques. The results for six 15-km × 15-km areas of interest in northern Minnesota, USA, indicate that areal estimates of tree volume, basal area, and density obtained from pixel predictions are comparable to plot-based estimates and estimates by conifer and deciduous classes are also comparable to plot-based estimates. When a mixed conifer/deciduous class was included, predictions for the mixed and deciduous class were confused.     

Parametric, bootstrap, and jackknife variance estimators for the k-Nearest Neighbors technique with illustrations using forest inventory and satellite image data

Nearest neighbors techniques have been shown to be useful for estimating forest attributes, particularly when used with forest inventory and satellite image data. Published reports of positive results have been truly international in scope. However, for these techniques to be more useful, they must be able to contribute to scientific inference which, for sample-based methods, requires estimates of uncertainty in the form of variances or standard errors. Several parametric approaches to estimating uncertainty for nearest neighbors techniques have been proposed, but they are complex and computationally intensive. For this study, two resampling estimators, the bootstrap and the jackknife, were investigated and compared to a parametric estimator for estimating uncertainty using the k-Nearest Neighbors (k-NN) technique with forest inventory and Landsat data from Finland, Italy, and the USA. The technical objectives of the study were threefold: (1) to evaluate the assumptions underlying a parametric approach to estimating k-NN variances; (2) to assess the utility of the bootstrap and jackknife methods with respect to the quality of variance estimates, ease of implementation, and computational intensity; and (3) to investigate adaptation of resampling methods to accommodate cluster sampling. The general conclusions were that support was provided for the assumptions underlying the parametric approach, the parametric and resampling estimators produced comparable variance estimates, care must be taken to ensure that bootstrap resampling mimics the original sampling, and the bootstrap procedure is a viable approach to variance estimation for nearest neighbor techniques that use very small numbers of neighbors to calculate predictions.     

Classification of Amazonian primary rain forest vegetation using Landsat ETM+ satellite imagery

A lack of spatially and thematically accurate vegetation maps complicates conservation and management planning, as well as ecological research, in tropical rain forests. Remote sensing has considerable potential to provide such maps, but classification accuracy within primary rain forests has generally been inadequate for practical applications. Here we test how accurately floristically defined forest types in lowland tropical rain forests in Peruvian Amazonia can be recognized using remote sensing data (Landsat ETM+ satellite image and STRM elevation model). Floristic data and a vegetation classification with four forest classes were available for eight line transects, each 8 km long, located in an area of ca 800 km². We compared two sampling unit sizes (line transect subunits of 200 and 500 m) and several image feature combinations to analyze their suitability for image classification. Mantel tests were used to quantify how well the patterns in elevation and in the digital numbers of the satellite image correlated with the floristic patterns observed in the field. Most Mantel correlations were positive and highly significant. Linear discriminant analysis was used first to build a function that discriminates between forest classes in the eight field-verified transects on the basis of remotely sensed data, and then to classify those parts of the line transects and the satellite image that had not been visited in the field. Classification accuracy was quantified by 8-fold crossvalidation. Two of the tierra firme (non-inundated) forest types were combined because they were too often misclassified. The remaining three forest types (inundated forest, terrace forest and Pebas formation/intermediate tierra firme forest) could be separated using the 500-m sampling units with an overall classification accuracy of 85% and a Kappa coefficient of 0.62. For the 200-m sampling units, the classification accuracy was clearly lower (71%, Kappa 0.35). The forest classification will be used as habitat data to study wildlife habitat use in the same area. Our results show that remotely sensed data and relatively simple classification methods can be used to produce reasonably accurate forest type classifications, even in structurally homogeneous primary rain forests.     

Adapting a global stratified random sample for regional estimation of forest cover change derived from satellite imagery

A desirable feature of a global sampling design for estimating forest cover change based on satellite imagery is the ability to adapt the design to obtain precise regional estimates, where a region may be a country, state, province, or conservation area. A sampling design stratified by an auxiliary variable correlated with forest cover change has this adaptability. A global stratified random sample can be augmented by additional sample units within a region selected by the same stratified protocol and the resulting sample constitutes a stratified random sample of the region. Stratified sampling allows increasing the sample size in a region by a few to many additional sample units. The additional sample units can be effectively allocated to strata to reduce the standard errors of the regional estimates, even though these strata were not initially constructed for the objective of regional estimation. A complete coverage map of deforestation within the Brazilian Legal Amazon (BLA) is used as a population to evaluate precision of regional estimates obtained by augmenting a global stratified random sample. The standard errors of the regional estimates for the BLA and states within the BLA obtained from the augmented stratified design were generally smaller than those attained by simple random sampling and systematic sampling.     

Improved classification of conservation tillage adoption using high temporal and synthetic satellite imagery

Conservation tillage management has been advocated for carbon sequestration and soil quality preservation purposes. Past satellite image analyses have had difficulty in differentiating between no-till (NT) and minimal tillage (MT) conservation classes due to similarities in surface residues, and may have been restricted by the availability of cloud-free satellite imagery. This study hypothesized that the inclusion of high temporal data into the classification process would increase conservation tillage accuracy due to the added likelihood of capturing spectral changes in MT fields following a tillage disturbance. Classification accuracies were evaluated for Random Forest models based on 250-m and 500-m MODIS, 30-m Landsat, and 30-m synthetic reflectance values. Synthetic (30-m) data derived from the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) were evaluated because high frequency Landsat image sets are often unavailable within a cropping season due to cloud issues. Classification results from a five-date Landsat model were substantially better than those reported by previous classification tillage studies, with 94% total and ≥ 88% class producer's accuracies. Landsat-derived models based on individual image scenes (May through August) yielded poor MT classifications, but a monthly increase in accuracy illustrated the importance of temporal sampling for capturing regional tillage disturbance signatures. MODIS-based model accuracies (90% total; ≥ 82% class) were lower than in the five-date Landsat model, but were higher than previous image-based and survey-based tillage classification results. Almost all the STARFM prediction-based models had classification accuracies higher than, or comparable to, the MODIS-based results (> 90% total; ≥ 84% class) but the resulting model accuracies were dependent on the MODIS/Landsat base pairs used to generate the STARFM predictions. Also evident within the STARFM prediction-based models was the ability for high frequency data series to compensate for degraded synthetic spectral values when classifying field-based tillage. The decision to use MODIS or STARFM-based data within conservation tillage analysis is likely situation dependent. A MODIS-based approach requires little data processing and could be more efficient for large-area mapping; however a STARFM-based analysis might be more appropriate in mixed-pixel situations that could potentially compromise classification accuracy.     

Comparison of two types of forest disturbance using multitemporal Landsat TM/ETM+ imagery and field vegetation data

Forest disturbances influence many landscape processes, including changes in microclimate, hydrology, and soil erosion. We analyzed the spectral response and temporal progress of two types of disturbances of spruce forest (bark beetle outbreak and clear-cuts) in the central part of Šumava Mountains at the border between the Czech Republic and Germany, Central Europe. The bark beetle (Ips typographus [L.]) outbreak in this region in the last 20 years resulted in regional-scale spruce forest decay. Clear-cutting was done here to prevent further bark-beetle propagation in the buffer zones.The aim of the study is to identify the differences in spectral response between the two types of forest disturbances and their temporal dynamics. General trends were analyzed throughout the study area, with sampled disturbance areas selected to assess the relationship between field vegetation data and their spectral response. Thirteen Landsat TM/ETM+ scenes from 1985 to 2007 were used for the assessment. The following spectral indices were estimated: NDMI, Tasseled Cap (Brightness, Greenness, Wetness), DI, and DI′. The DI′, Wetness, and Brightness indices show the highest sensitivity to forest disturbance for both disturbance types (clear-cuts and bark beetle outbreak). The multitemporal analysis distinguished three different stages of development. The highest spectral differences between the clear-cuts and the bark beetle disturbances were found in the period between 1996 and 2004 with increased levels of forest disturbance (repeated measures ANOVA, Scheffé post hoc test; p ≤ 0.05). Clear-cut disturbance resulted in significantly higher spectral differences from the original forest and occurred as a more discrete event in comparison to bark beetle outbreak.     

Design-based approach to k-nearest neighbours technique for coupling field and remotely sensed data in forest surveys

The statistical properties of the k-NN estimators are investigated in a design-based framework, avoiding any assumption about the population under study. The issue of coupling remotely sensed digital imagery with data arising from forest inventories conducted using probabilistic sampling schemes is considered. General results are obtained for the k-NN estimator at the pixel level. When averages (or totals) of forest attributes for the whole study area or sub-areas are of interest, the use of the empirical difference estimator is proposed. The estimator is shown to be approximately unbiased with a variance admitting unbiased or conservative estimators. The performance of the empirical difference estimator is evaluated by an extensive simulation study performed on several populations whose dimensions and covariate values are taken from a real case study. Samples are selected from the populations by means of simple random sampling without replacement. Comparisons with the generalized regression estimator and Horvitz-Thompson estimators are also performed. An application to a local forest inventory on a test area of central Italy is considered.     

Model-based mean square error estimators for k-nearest neighbour predictions and applications using remotely sensed data for forest inventories

New model-based estimators of the uncertainty of pixel-level and areal k-nearest neighbour (k_nn) predictions of attribute Y from remotely-sensed ancillary data X are presented. Non-parametric functions predict Y from scalar ‘Single Index Model’ transformations of X. Variance functions generated estimates of the variance of Y. Three case studies, with data from the Forest Inventory and Analysis program of the U.S. Forest Service, the Finnish National Forest Inventory, and Landsat ETM+ ancillary data, demonstrate applications of the proposed estimators. Nearly unbiased k_nn predictions of three forest attributes were obtained. Estimates of mean square error indicate that k_nn is an attractive technique for integrating remotely-sensed and ground data for the provision of forest attribute maps and areal predictions.     

Estimating aboveground carbon in a catchment of the Siberian forest tundra: Combining satellite imagery and field inventory

This study was part of an interdisciplinary research project on soil carbon and phytomass dynamics of boreal and arctic permafrost landscapes. The 45 ha study area was a catchment located in the forest tundra in northern Siberia, approximately 100 km north of the Arctic Circle.The objective of this study was to estimate aboveground carbon (AGC) and assess and model its spatial variability. We combined multi-spectral high resolution remote sensing imagery and sample based field inventory data by means of the k-nearest neighbor (k-NN) technique and linear regression.Field data was collected by stratified systematic sampling in August 2006 with a total sample size of n = 31 circular nested sample plots of 154 m² for trees and shrubs and 1 m² for ground vegetation. Destructive biomass samples were taken on a sub-sample for fresh weight and moisture content. Species-specific allometric biomass models were constructed to predict dry biomass from diameter at breast height (dbh) for trees and from elliptic projection areas for shrubs.Quickbird data (standard imagery product), acquired shortly before the field campaign and archived ASTER data (Level-1B product) of 2001 were geo-referenced, converted to calibrated radiances at sensor and used as carrier data. Spectral information of the pixels which were located in the inventory plots were extracted and analyzed as reference set. Stepwise multiple linear regression was applied to identify suitable predictors from the set of variables of the original satellite bands, vegetation indices and texture metrics. To produce thematic carbon maps, carbon values were predicted for all pixels of the investigated satellite scenes. For this prediction, we compared the kNN distance-weighted classifier and multiple linear regression with respect to their predictions.The estimated mean value of aboveground carbon from stratified sampling in the field is 15.3 t/ha (standard error SE = 1.50 t/ha, SE% = 9.8%). Zonal prediction from the k-NN method for the Quickbird image as carrier is 14.7 t/ha with a root mean square error RMSE = 6.42 t/ha, RMSE_r = 44%) resulting from leave-one-out cross-validation. The k-NN-approach allows mapping and analysis of the spatial variability of AGC. The results show high spatial variability with AGC predictions ranging from 4.3 t/ha to 28.8 t/ha, reflecting the highly heterogeneous conditions in those permafrost-influenced landscapes. The means and totals of linear regression and k-NN predictions revealed only small differences but some regional distinctions were recognized in the maps.     

Glacial cover mapping (1987-1996) of the Cordillera Blanca (Peru) using satellite imagery

Multitemporal glacier area mapping is a key element in accurately determining fresh water reserves, as well as providing an indicator of climate change.In Peru, the first glacier inventory was based on visual interpretation of aerial photos, requiring several years of effort. Landsat Thematic Mapper satellite imagery, on the other hand, provides an increasingly employed alternative for the monitoring of changes in glacier area and in other glaciological parameters.By means of Normalized Difference Snow Index (NDSI) computations on TM images, an estimate of the glacierized area in Cordillera Blanca (Peru) was carried out for 1987 (643±63 km²) and 1996 (600±61 km²). Compared to an estimate of 721 km² in 1970, it can be concluded that the glacier area has retreated in this massif by more than 15% in 25 years.     

Using satellite imagery and the k-nearest neighbors technique as a bridge between strategic and management forest inventories

Strategic forest inventory programs produce forest resource estimates for large areas such as states and provinces using data collected for a large number of variables on a relatively sparse array of field plots. Management inventories produce stand-level estimates to guide management decisions using data obtained with sampling intensities much greater than for strategic inventories. The costs associated with these greater sampling intensities have motivated investigations of alternatives to traditional sample-based management inventories. This study focused on a relatively inexpensive alternative to management inventories that uses strategic forest inventory plot data, Landsat Thematic Mapper (TM) satellite imagery, and the k-Nearest Neighbors (k-NN) technique. The approach entailed constructing stem density and basal area per unit area maps from which stand-level means were estimated as averages of k-NN pixel predictions. The study included investigations of the benefits of selecting optimal combinations of k-NN feature space variables derived from the TM imagery and the benefits of modifying the k-NN technique to eliminate spurious nearest neighbors. For both the stem density and basal area per unit area training data, the selection of optimal feature space covariates produced less than 1.5% improvement in root mean square error relative to using all covariates. The k-NN modification improved the sum of mean squared deviations for stand-level stem density and basal area per unit area estimates by 7–20% depending on the k-NN feature space covariates. For the best combination of feature space covariates, estimates of stand-level means were within confidence intervals for validation estimates for 11 of 12 stands for stem density and for 10 of 12 stands for basal area per unit area.     

Probability- and model-based approaches to inference for proportion forest using satellite imagery as ancillary data

Estimates of forest area are among the most common and useful information provided by national forest inventories. The estimates are used for local and national purposes and for reporting to international agreements such as the Montréal Process, the Ministerial Conference on the Protection of Forests in Europe, and the Kyoto Protocol. The estimates are usually based on sample plot data and are calculated using probability-based estimators. These estimators are familiar, generally unbiased, and entail only limited computational complexity, but they do not produce the maps that users are increasingly requesting, and they generally do not produce sufficiently precise estimates for small areas. Model-based estimators overcome these disadvantages, but they may be biased and estimation of variances may be computationally intensive. The study objective was to compare probability- and model-based estimators of mean proportion forest using maps based on a logistic regression model, forest inventory data, and Landsat imagery. For model-based estimators, methods for evaluating bias and reducing the computational intensity were also investigated. Four conclusions were drawn: the logistic regression model exhibited no serious lack of fit to the data; all the estimators produced comparable estimates for mean proportion forest, except for small areas; probability-based inferences enhanced using maps produced increased precision; and the computational intensity associated with estimating variances for model-based estimators can be greatly reduced with no detrimental effects.     

Assessment of deforestation in the Lower Amazon floodplain using historical Landsat MSS/TM imagery

The floodplain forests bordering the Amazon River have outstanding ecological, economic, and social importance for the region. However, the original distribution of these forests is not well known, since they have suffered severe degradation since the 16th century. The previously published vegetation map of the Amazon River floodplain (Hess et al., 2003), based on data acquired in 1996, shows enormous difference in vegetation cover classes between the regions upstream and downstream of the city of Manaus. The upper floodplain is mostly covered by forests, while the lower floodplain is predominantly occupied by grasses and shrubs.This study assesses deforestation in the Lower Amazon floodplain over a ~ 30 year period by producing and comparing a historical vegetation map based on MSS/Landsat images acquired in the late 1970s with a recent vegetation map produced from TM/Landsat images obtained in 2008. The maps were generated through the following steps: 1) normalization and mosaicking of images for each decade; 2) application of a linear mixing model transformation to produce vegetation, soil and shade fraction-images; and 3) object-oriented image analysis and classification. For both maps, the following classes were mapped: floodplain forest, non-forest floodplain vegetation, bare soil and open water. The two maps were combined using object-level Boolean operations to identify time transitions among the mapped classes, resulting in a map of the land cover change occurred over ~ 30 years. Ground information collected at 168 ground points was used to build confusion matrices and calculate Kappa indices of agreement. A survey strategy combining field observations and interviews allowed the collection of information about both recent and historical land cover for validation purposes. Kappa values (0.77, 0.75 and 0.75) indicated the good quality of the maps, and the error estimates were used to adjust the estimated deforested area to a value of 3457 km² ± 1062 km² (95% CI) of floodplain deforestation over the ~ 30 years.     

Non-parametric and parametric methods using satellite images for estimating growing stock volume in alpine and Mediterranean forest ecosystems

This paper describes applications of non-parametric and parametric methods for estimating forest growing stock volume using Landsat images on the basis of data measured in the field, integrated with ancillary information. Several k-Nearest Neighbors (k-NN) algorithm configurations were tested in two study areas in Italy belonging to Mediterranean and Alpine ecosystems. Field data were acquired by the regional forest inventory and forest management plans, and satellite images are from Landsat 5 TM and Landsat 7 ETM+. The paper describes the data used, the methodologies adopted and the results achieved in terms of pixel level accuracy of forest growing stock volume estimates. The results show that several factors affect estimation accuracy when using the k-NN method. For the two test areas a total of 3500 different configurations of the k-NN algorithm were systematically tested by changing the number and type of spectral and ancillary input variables, type of multidimensional distance measures, number of nearest neighbors and methods for spectral feature extraction using the leave-one-out (LOO) procedure. The best k-NN configurations were then used for pixel level estimation; the accuracy was estimated with a bootstrapping procedure; and the results were compared to estimates obtained using parametric regression methods implemented on the same data set.

The best k-NN growing stock volume pixel level estimates in the Alpine area have a Root Mean Square Error (RMSE) ranging between 74 and 96 m³ ha^− 1 (respectively, 22% and 28% of the mean measured value) and between 106 and 135 m³ ha^− 1 (respectively, 44% and 63% of the mean measured value) in the Mediterranean area. On the whole, the results cast a promising light on the use of non-parametric techniques for forest attribute estimation and mapping with accuracy high enough to support forest planning activities in such complex landscapes. The results of the LOO analyses also highlight the importance of a local empirical optimization phase of the k-NN procedure before defining the best algorithm configuration. In the tests performed the pixel level accuracy increased, depending on the k-NN configuration, as much as 100%.     

Assessment of fire severity and species diversity in the southern Appalachians using Landsat TM and ETM+ imagery

Relatively little is known about the disturbance ecology of large wildfires in the southern Appalachians. The occurrence of a 4000-ha wildfire in the Linville Gorge Wilderness area in western North Carolina has provided a rare opportunity to study a large fire with a range of severities. The objectives of this study were to 1) assess the potential for using multi-temporal Landsat imagery to map fire severity in the southern Appalachians, 2) examine the influences of topography and forest community type on the spatial pattern of fire severity; and 3) examine the relationship between predicted fire severity and changes in species richness. A non-linear regression equation predicted a field-based composite burn index (CBI) as a function of change in the Normalized Burn Ratio (dNBR) with an R² of 0.71. Fire severity was highest on drier landforms located on upper hillslopes, ridges, and on southwest aspects, and was higher in pine communities than in other forest types. Predicted CBI was positively correlated with changes in species richness and with the post-fire cover of pine seedlings (Pinus virginiana, P. rigida, and P. pungens), suggesting that burn severity maps can be used to predict community-level fire effects across large landscapes. Despite the relatively large size of this fire for the southern Appalachians, severity was strongly linked to topographic variability and pre-fire vegetation, and spatial variation in fire severity was correlated with changes in species richness. Thus, the Linville Gorge fire appears to have generally reinforced the ecological constraints imposed by underlying environmental gradients.     

Using the spatial and spectral precision of satellite imagery to predict wildlife occurrence patterns

We investigated the potential of using unclassified spectral data for predicting the distribution of three bird species over a ∼400,000 ha region of Michigan's Upper Peninsula using Landsat ETM+ imagery and 433 locations sampled for birds through point count surveys. These species, Black-throated Green Warbler, Nashville Warbler, and Ovenbird, were known to be associated with forest understory features during breeding. We examined the influences of varying two spatially explicit classification parameters on prediction accuracy: 1) the window size used to average spectral values in signature creation and 2) the threshold distance required for bird detections to be counted as present. Two accuracy measurements, proportion correctly classified (PCC) and Kappa, of maps predicting species' occurrences were calculated with ground data not used during classification. Maps were validated for all three species with Kappa values > 0.3 and PCC > 0.6. However, PCC provided little information other than a summary of sample plot frequencies used to classify species' presence and absence. Comparisons with rule-based maps created using the approach of Gap Analysis showed that spectral information predicted the occurrence of these species that use forest subcanopy components better than could be done using known land cover associations (Kappa values 0.1 to 0.3 higher than Gap Analysis maps). Accuracy statistics for each species were affected in different ways by the detection distance of point count surveys used to stratify plots into presence and absence classes. Moderate-to-large detection distances (100 m and 180 m) best classified maps of Black-throated Green Warbler and Nashville Warbler occurrences, while moderate detection distances (50 m and 100 m), which ignored remote observations, provided the best source of information for classification of Ovenbird occurrence. Window sizes used in signature creation also influenced accuracy statistics but to a lesser extent. Highest Kappa values of majority maps were typically obtained using moderate window sizes of 9 to 13 pixels (0.8 to 1.2 ha), which are representative of the study species territory sizes. The accuracy of wildlife occurrence maps classified from spectral data will therefore differ given the species of interest, the spatial precision of occurrence records used as ground references and the number of pixels included in spectral signatures. For these reasons, a quantitative examination is warranted to determine how subjective decisions made during image classifications affect prediction accuracies.     

Application of high spatial resolution satellite imagery for riparian and forest ecosystem classification

Terrestrial Ecosystem Mapping provides critical information to land and resource managers by incorporating information on climate, physiography, surficial material, soil, and vegetation structure. The main objective of this research was to determine the capacity of high spatial resolution satellite image data to discriminate vegetation structural stages in riparian and adjacent forested ecosystems as defined using the British Columbia Terrestrial Ecosystem Mapping (TEM) scheme. A high spatial resolution QuickBird image, captured in June 2005, and coincident field data covering the riparian area of Lost Shoe Creek and adjacent forests on Vancouver Island, British Columbia, was used in this analysis. Semi-variograms were calculated to assess the separability of vegetation structural stages and assess which spatial scales were most appropriate for calculation of grey-level co-occurrence texture measures to maximize structural class separation. The degree of spatial autocorrelation showed that most vegetation structural types in the TEM scheme could be differentiated and that window sizes of 3 × 3 pixels and 11 × 11 pixels were most appropriate for image texture calculations. Using these window sizes, the texture analysis showed that co-occurrence contrast, dissimilarity, and homogeneity texture measures, based on the bands in the visible part of the spectrum, provided the most significant statistical differentiation between vegetation structural classes. Subsequently, an object-oriented classification algorithm was applied to spectral and textural transformations of the QuickBird image data to map the vegetation structural classes. Using both spectral and textural image bands yielded the highest classification accuracy (overall accuracy = 78.95%). The inclusion of image texture increased the classification accuracies of vegetation structure by 2-19%. The results show that information on vegetation structure can be mapped effectively from high spatial resolution satellite image data, providing an additional tool to ongoing aerial photograph interpretation.     

An automated approach for segmenting and classifying a large sample of multi-date Landsat imagery for pan-tropical forest monitoring

The TREES-3 project of the Joint Research Centre aims at assessing tropical forest cover changes for the periods 1990-2000 and 2000-2010 using a sample-based approach. This paper refers to the 1990-2000 assessment. Extracts of Landsat satellite imagery (20 km × 20 km) are analyzed for these reference dates for more than 4000 sample sites distributed systematically across the tropical belt. For the processing and analysis of such a large amount of satellite imagery a robust methodological approach for forest mapping and change detection has been developed. This approach comprises two automated steps of multi-date image segmentation and object-based land cover classification (based on a supervised spectral library), followed by an intense phase of visual control and expert refinement. Image segmentation is done at two spatial scales, introducing the concept of a minimum mapping unit via the automated selection of a site-specific scale parameter. The automated segmentation of land cover polygons and the pre-classification of land cover types mainly aim at avoiding manual delineation and at reducing the efforts of visual interpretation of land cover to a reasonable level, making the analysis of 4000 sample sites feasible. The importance of visual control and correction can be perceived when comparing to the initial automatic classification result: about 20% of the polygon labels were changed through expert knowledge by visual interpretation. The component of visual refinement of the mapping results had also a notable impact on forest area and change estimates: for a set of sample sites in Southeast Asia (~ 90% of all sites of SE-Asia) the rate of change in tree cover (deforestation) was assessed at 0.9% and 1.6% before and after visual control, respectively.     

Cloud detection in Landsat imagery of ice sheets using shadow matching technique and automatic normalized difference snow index threshold value decision

This work presents a new algorithm designed to detect clouds in satellite visible and infrared (IR) imagery of ice sheets. The approach identifies possible cloud pixels through the use of the normalized difference snow index (NDSI). Possible cloud pixels are grown into regions and edges are determined. Possible cloud edges are then matched with possible cloud shadow regions using knowledge of the solar illumination azimuth. A scoring index quantifies the quality of each match resulting in a classified image. The best value of the NDSI threshold is shown to vary significantly, forcing the algorithm to be iterated through many threshold values. Computational efficiency is achieved by using sub-sampled images with only minor degradation in cloud-detection performance. The algorithm detects all clouds in each of eight test Landsat-7 images and makes no incorrect cloud classifications.     

Estimating stand structural details using nearest neighbor analyses to link ground data, forest cover maps, and Landsat imagery

Forest information over a landscape is often represented as a spatial mosaic of polygons, separated by differences in species composition, height, age, crown closure, productivity, and other variables. These polygons are commonly delineated on medium-scale photography (e.g., 1:15,000) by a photo-interpreter familiar with the inventory area, and displayed and stored in a Geographic Information System (GIS) layer as a forest cover map. Forest cover maps are used for multiple purposes including timber and habitat supply analyses, and carbon inventories, at a regional or management unit level, and for parks planning, operational planning, and selection of stands for many purposes at a local level. Attribute data for each polygon commonly include the variables used to delineate the polygon, and other variables that can be measured or estimated using these medium-scale photographs. Additional measures that can only be obtained via expensive ground measures or possibly on high resolution photographs (e.g., volume per unit area, biomass components per unit area, tree-list of species and diameters) are available only for a sample of polygons, or may have been gathered independently using a sample survey over the land area. Improved linkages over a variety of data sources may help to support landscape level analyses. This study presents an approach to combine information from a systematic (grid) ground survey, forest cover (polygon) data, and Landsat Thematic Mapper (TM) imagery using variable-space nearest neighbor methods to estimate (i) mean ground-measured attributes for each polygon, in particular, volume per ha (m³/ha), stems per ha, and quadratic mean diameter for each polygon; and (ii) variation of these ground attributes within polygons. The approach was initially evaluated using Monte Carlo simulations with known measures of these attributes. Nearest neighbor methods were then applied to an approximate 5000 ha area (about 1000 polygons) of high productivity, mountainous forests located near the Pacific Coast of British Columbia, Canada. Based on the simulation results, the use of Landsat pixel reflectances to estimate volume per ha, average tree size (i.e., quadratic mean diameter), and stems per ha did not show great promise in improving estimates for each polygon over using forest cover data alone. However, in application, the use of remotely sensed data provided estimates of within-polygon variability. At the same time, the estimated means of these three imputed variables over the entire study area were very similar to the representative sample estimates using the ground data only. Extensions to other variables such as ranges of diameters and numbers of snags may also be possible providing useful data for habitat and forest growth analysis.