The use of prior probabilities in maximum likelihood classification of remotely sensed data

The expected distribution of classes in a final classification map can be used to improve classification accuracies. Prior information is incorporated through the use of prior probabilities—that is, probabilities of occurrence of classes which are based on separate, independent knowledge concerning the area to be classified. The use of prior probabilities in a classification system is sufficiently versatile to allow (1) prior weighting of output classes based on their anticipated sizes; (2) the merging of continuously varying measurements (multispectral signatures) with discrete collateral information datasets (e.g., rock type, soil type); and (3) the construction of time-sequential classification systems in which an earlier classification modifies the outcome of a later one. The prior probabilities are incorporated by modifying the maximum likelihood decision rule employed in a Bayesian-type classifier to calculate a posteriori probabilities of class membership which are based not only on the resemblance of a pixel to the class signature, but also on the weight of the class which is estimated for the final output classification. In the merging of discrete collateral information with continuous spectral values into a single classification, a set of prior probabilities (weights) is estimated for each value which the discrete collateral variable may assume (e.g., each rock type or soil type). When maximum likelihood calculations are performed, the prior probabilities appropriate to the particular pixel are used in classification. For time-sequential classification, the prior classification of a pixel indexes a set of appropriate conditional probabilities reflecting either the confidence of the investigator in the prior classification or the extent to which the prior class identified is likely to change during the time period of interest.     

Incorporating domain knowledge and spatial relationships into land cover classifications: a rule‐based approach

For some tropical regions, remote sensing of land cover yields unacceptable results, particularly as the number of land cover classes increases. This research explores the utility of incorporating domain knowledge and multiple algorithms into land cover classifications via a rule‐based algorithm for a series of satellite images. The proposed technique integrates the fundamental, knowledge‐based interpretation elements of remote sensing without sacrificing the ease and consistency of automated, algorithm‐based processing. Compared with results from a traditional maximum likelihood algorithm, classification accuracy was improved substantially for each of the six land cover classes and all three years in the image series. Use of domain knowledge proved effective in accurately classifying problematic tropical land covers, such as tropical deciduous forest and seasonal wetlands. Results also suggest that ancillary data may be most useful in the classification of historic images, where the greatest improvement was observed relative to results from maximum likelihood. The cost of incorporating contextual knowledge and extensive spatial data sets may be justified, since results from the proposed technique suggest a considerable improvement in accuracy may be achieved.     

Stand delineation and composition estimation using semi-automated individual tree crown analysis

Stand delineation and species composition estimation are cornerstones of forest inventory mapping and key elements to forest management decision making. Improved mapping techniques are constantly being sought in terms of speed, consistency, accuracy, level of detail, and overall effectiveness. Semi-automated analysis of high-resolution imagery at the individual tree crown level may offer such benefits. Methods, however, need to be developed and tested under a variety of forest conditions.High-resolution (60 cm) multispectral airborne imagery was acquired over a predominantly young conifer forest and plantation test area on the west coast of Canada. Automated tree isolation algorithms were applied to the data in order to delineate tree crowns or clusters of crowns. An object-oriented single tree classification was conducted using a maximum likelihood classifier. Stands of similar species composition, closure, and stem density were defined through a sequence that first generated images of these parameters from the automated delineation and classification, used these as input to an unsupervised classification, and then filtered and smoothed the resulting classification clusters. Because of the dense nature of the stands and small crowns on the site, the isolation process often delineated clusters of several trees. Species classification accuracy was determined by comparing the average stand composition from the automated technique to that derived from ground transects or plots. Species classification was good, with average composition error (difference between field measured and automated composition) over all 16 test stands being 7.25%. Most errors for individual species in stands were below 20%, but a few were up to 30%. The automatically generated stand boundaries mimicked well those of known plantation and interpreted inventory boundaries. The automated technique created a few larger stands and some additional small stands in areas of complex forest structure. Overall, for the young fairly uniform stands of the site, both stand delineation and species composition estimation were of a quality suitable for operational use in inventory and forest management. Further development and testing is needed to extend results to situations covering large areas, multiple flight lines, varied topography, and different forest conditions.     

An evaluation of fuzzy and texture-based classification approaches for mapping regenerating tropical forest classes from Landsat-TM data

Two classification approaches were investigated for the mapping of tropical forests from Landsat-TM data of a region north of Manaus in the Brazilian state of Amazonas. These incorporated textural information and made use of fuzzy approaches to classification. In eleven class classifications the texture-based classifiers (based on a Markov random field model) consistently provided higher classification accuracies than conventional per-pixel maximum likelihood and minimum distance classifications, indicating that they are more able to characterize accurately several regenerating forest classes. Measures of the strength of class memberships derived from three classification algorithms (based on the probability density function, a posteriori probability and the Mahalanobis distance) could be used to derive fuzzy image classifications and be used in post-classification processing. The latter, involving either the summation of class memberships over a local neighbourhood or the application of homogeneity measures, were found to increase classification accuracy by some 10 per cent in comparison with a conventional maximum likelihood classification, a result of comparable accuracy to that derived from the texture-based classifications.     

Improving satellite images classification using remote and ground data integration by means of stochastic simulation

A methodology is proposed, to assess land surface cover classification using a geostatistical methodology of stochastic simulation, direct sequential cosimulation, to combine field observations with remotely sensed data classified with the classical algorithm of maximum likelihood classification. This procedure has two main advantages: (1) incorporation of a spatial continuity statistics; and (2) integration of different scales of information, contained in polygons (training areas) and point information (field observations), which also involves different qualities of information that is less reliable and more reliable, respectively. Moreover, this methodology allows production not only of a classified map, but also of maps of occupation proportions and of uncertainty for each thematic class. Local co‐regionalization models are applied to account for local differences in both field data availability and distribution, and the correlation between these hard data and the classified satellite images as soft data. The methodology is based on two criteria: the influence of the hard data dependent on their availability and proportional to their proximity; and the influence of the soft data dependent on their local correlation to the hard data. The method is applied to a study of four economically important forest tree species on the Setúbal Peninsula (south of Lisbon, Portugal). The results show more contiguous forest covers, i.e. more spatial contiguity, than the classical classification. In comparison to a contemporary field inventory, the proposed method improved forest cover estimations, showing a difference of only 3%.     

Transductive methods for the distributed ensemble classification problem

We consider ensemble classification for the case where there is no common labeled training data for jointly designing the individual classifiers and the function that aggregates their decisions. This problem, which we call distributed ensemble classification, applies when individual classifiers operate (perhaps remotely) on different sensing modalities and when combining proprietary or legacy classifiers. The conventional wisdom in this case is to apply fixed rules of combination such as voting methods or rules for aggregating probabilities. Alternatively, we take a transductive approach, optimizing the combining rule for an objective function measured on the unlabeled batch of test data. We propose maximum likelihood (ML) objectives that are shown to yield well-known forms of probabilistic aggregation, albeit with iterative, expectation-maximization-based adjustment to account for mismatch between class priors used by individual classifiers and those reflected in the new data batch. These methods are extensions, for the ensemble case, of the work of Saerens, Latinne, and Decaestecker (2002). We also propose an information-theoretic method that generally outperforms the ML methods, better handles classifier redundancies, and addresses some scenarios where the ML methods are not applicable. This method also well handles the case of missing classes in the test batch. On UC Irvine benchmark data, all our methods give improvements in classification accuracy over the use of fixed rules when there is prior mismatch.     

Comparison of Landsat image classification methods for detecting mangrove forests in Sundarbans

Remote-sensing images taken from the Landsat Enhanced Thematic Mapper Plus (ETM+) sensor with a spatial resolution of 30 m were applied for mapping and inventory of mangrove forest areas in Sundarbans, on both sides of the border between Bangladesh and India. Three different classification methods – unsupervised classification with k-means clustering, supervised classification using the maximum likelihood decision rule, and band-ratio supervised classification – were tested and compared in terms of the top of the atmosphere reflectance images. Spectral signature and principal component analyses were applied to select the appropriate band combinations prior to the band ratio–supervised classification. Our results show that the band ratio method is superior to the unsupervised or supervised classification methods considering the visual inspection, producer's and user's accuracy, as well as the overall accuracy of the all the classes in the image. The best discrimination of mangrove/nonmangrove boundary can be achieved when the combinations of B4/B2 (band 4/band 2), B5/B7, and B7/B4 are employed from the ETM+?bands.     

Floodplain roughness parameterization using airborne laser scanning and spectral remote sensing

Floodplain roughness parameterization is one of the key elements of hydrodynamic modeling of river flow, which is directly linked to exceedance levels of the embankments of lowland fluvial areas. The present way of roughness mapping is based on manually delineated floodplain vegetation types, schematized as cylindrical elements of which the height (m) and the vertical density (the projected plant area in the direction of the flow per unit volume, m^− 1) have to be assigned using a lookup table. This paper presents a novel method of automated roughness parameterization. It delivers a spatially distributed roughness parameterization in an entire floodplain by fusion of CASI multispectral data with airborne laser scanning (ALS) data. The method consists of three stages: (1) pre-processing of the raw data, (2) image segmentation of the fused data set and classification into the dominant land cover classes (KHAT = 0.78), (3) determination of hydrodynamic roughness characteristics for each land cover class separately. In stage three, a lookup table provides numerical values that enable roughness calculation for the classes water, sand, paved area, meadows and built-up area. For forest and herbaceous vegetation, ALS data enable spatially detailed analysis of vegetation height and density. The hydrodynamic vegetation density of forest is mapped using a calibrated regression model. Herbaceous vegetation cover is further subdivided in single trees and non-woody vegetation. Single trees were delineated using a novel iterative cluster merging method, and their height is predicted (R² = 0.41, rse = 0.84 m). The vegetation density of single trees was determined in an identical way as for forest. Vegetation height and density of non-woody herbaceous vegetation were also determined using calibrated regression models. A 2D hydrodynamic model was applied with the results of this novel method, and compared with a traditional roughness parameterization approach. The modeling results showed that the new method is well able to provide accurate output data. The new method provides a faster, repeatable, and more accurate way of obtaining floodplain roughness, which enables regular updating of river flow models.     

The effects of rectification and Global Positioning System errors on satellite image-based estimates of forest area

Satellite image-based maps of forest attributes are of considerable interest and are used for multiple purposes such as international reporting by countries that have no national forest inventory and small area estimation for all countries. Construction of the maps typically entails, in part, rectifying the satellite images to a geographic coordinate system, observing ground plots whose coordinates are obtained from Global Positioning System (GPS) receivers that are calibrated to the same geographic coordinate system, and then matching ground plots to image pixels containing the centers of the ground plots. Errors in rectification and GPS coordinates cause observations of ground attributes to be associated with spectral values of incorrect pixels which, in turn, introduces classification errors into the resulting maps. The most important finding of the study is that for common magnitudes of rectification and GPS errors, as many as half the ground plots may be assigned to incorrect pixels. The effects on areal estimates obtained by aggregating class predictions for individual pixels are deviation of the estimates from their true values, erroneous confidence intervals, and incorrect inferences. Results are reported in detail for both probability-based (design-based) and model-based approaches to inference for proportion forest area using maps constructed from Landsat imagery, forest inventory plot observations and a logistic regression model.     

Forest fire analysis with remote sensing data

Forest fires cause major damage to the environment, human health and property, and endanger life. Fires can be monitored and analysed over large areas in a timely and cost-effective manner by using satellite sensor imagery in combination with spatial analysis as provided by Geographical Information Systems (GIS). In this study, the forest area damage caused by a large fire which occurred in the Marmaris, province of Mugla in July 1996 was analysed using satellite sensor images. Digital image processing methods, such as spectral profile analysis, vegetation indices and multispectral classification, were applied to the satellite sensor images acquired before and after the forest fire. Besides the conventional maximum likelihood classification algorithm, a multilayer feed-forward neural network architecture was also used for comparison and evaluation of its effectiveness. A GIS database was constructed from the raster (satellite sensor data), vector (the forest type and topographical maps) and ancillary data (meteorological data). The GIS is being used to develop an information and decision support system to monitor and predict forest fire activity, and to enhance fire management efficiency. This study highlights the deficiencies in the current approach to fire management and emphasizes the need for an improved method along the lines outlined.     

The role of topographic correction in mapping recently burned Mediterranean forest areas from LANDSAT TM images

Operational use of remote sensing as a tool for post‐fire, Mediterranean forest management has been limited by problems of classification accuracy arising from confusion of burned and non‐burned areas. Frequently, this occurs as a result of slope illumination and shadowing effects caused by the complex topography encountered in many forested areas. Cloud shadows can also be a problem. The aim of this work was to investigate how image classification results could be improved by removing the illumination effects of topography from satellite images. This was achieved by applying supervised classification to both uncorrected and topographically corrected LANDSAT TM data for a site on the Greek island of Thasos. The classification methodology included atmospheric and geometric correction, field‐based training, seperability/contingency analysis and maximum likelihood processing. The classification scheme was determined on the basis of consultation with the Greek Forest Service. Overlay of the resulting class maps enabled comparison of the total burned area and its spatial extent using the two different approaches to processing. The results of each approach were compared with the forest perimeter map generated by the Forest Service using traditional survey methods. Accuracy assessment and error analysis clearly indicated that the removal of the topographic effect from the satellite image before its classification resulted in more accurate mapping of the burned area. It is concluded that operational use of satellite remote sensing for forest fire management depends on accurate, robust, widely available and proven techniques. Topographic correction should now be regarded as an essential element of any classification methodology which will be used for operational, post‐fire management of forests in complex Mediterranean landscapes.     

Mapping land use/cover distribution on a mountainous tropical island using remote sensing and GIS

The land use/cover distribution on Langkawi Island, Malaysia was mapped using remote sensing and a Geographic Information System (GIS). A Landsat Thematic Mapper (TM) satellite image taken in March 1995 was processed, geocorrected and analysed using IDRISI, raster-based GIS software. An unsupervised classification was performed based on spectral data from a composite image of the bands TM3, TM4 and TM5. Using this output, field data together with available secondary data consisting of topography, land use and soil maps were used to perform a maximum likelihood supervised classification. The overall accuracy of the output image was 90% and individual class accuracy ranged from 74% for rubber to 100% for paddy fields. The classified areas on the image were mainly confined to the mountainous and hilly regions on the island. A shaded relief map, simulating sunshine conditions, showed that the unclassified areas are located in the shadowed slopes, i.e. the slopes facing west. Consequently, the imagery was subdivided on the basis of slope aspect and a stratified classification was performed. As a result of this procedure, the overall accuracy increased to 92% and the individual class accuracy for the inland forest class increased by 9% to 90% . Using IDRISI, individual class areas as well as percentages were calculated. The kappa coefficient for the classified image was 0.90. Qualitative analysis indicates that topography is the main control on the spatial distribution of land use/cover types on the island. As Langkawi Island has been developing rapidly over the last decade, successful planning will require reliable information about land use/cover distribution and change. This study illustrates that remote sensing and GIS techniques are capable of providing such information.     

Objective priors from maximum entropy in data classification

Lack of knowledge of the prior distribution in classification problems that operate on small data sets may make the application of Bayes’ rule questionable. Uniform or arbitrary priors may provide classification answers that, even in simple examples, may end up contradicting our common sense about the problem. Entropic priors (EPs), via application of the maximum entropy (ME) principle, seem to provide good objective answers in practical cases leading to more conservative Bayesian inferences. EP are derived and applied to classification tasks when only the likelihood functions are available. In this paper, when inference is based only on one sample, we review the use of the EP also in comparison to priors that are obtained from maximization of the mutual information between observations and classes. This last criterion coincides with the maximization of the KL divergence between posteriors and priors that for large sample sets leads to the well-known reference (or Bernardo’s) priors. Our comparison on single samples considers both approaches in prospective and clarifies differences and potentials. A combinatorial justification for EP, inspired by Wallis’ combinatorial argument for entropy definition, is also included.The application of the EP to sequences (multiple samples) that may be affected by excessive domination of the class with the maximum entropy is also considered with a solution that guarantees posterior consistency. An explicit iterative algorithm is proposed for EP determination solely from knowledge of the likelihood functions. Simulations that compare EP with uniform priors on short sequences are also included.     

Application of spectral mixture analysis to Amazonian land-use and land-cover classification

Abundant vegetation species and associated complex forest stand structures in moist tropical regions often create difficulties in accurately classifying land-use and land-cover (LULC) features. This paper examines the value of spectral mixture analysis (SMA) using Landsat Thematic Mapper (TM) data for improving LULC classification accuracy in a moist tropical area in Rondônia, Brazil. Different routines, such as constrained and unconstrained least-squares solutions, different numbers of endmembers, and minimum noise fraction transformation, were examined while implementing the SMA approach. A maximum likelihood classifier was also used to classify fraction images into seven LULC classes: mature forest, intermediate secondary succession, initial secondary succession, pasture, agricultural land, water, and bare land. The results of this study indicate that reducing correlation between image bands and using four endmembers improve classification accuracy. The overall classification accuracy was 86.6% for the seven LULC classes using the best SMA processing routine, which represents very good results for such a complex environment. The overall classification accuracy using a maximum likelihood approach was 81.4%. Another finding is that use of constrained or unconstrained solutions for unmixing the atmospherically corrected or raw Landsat TM images does not have significant influence on LULC classification performances when image endmembers are used in a SMA approach.     

Automatic land cover analysis for Tenerife by supervised classification using remotely sensed data

Automatic land cover classification from satellite images is an important topic in many remote sensing applications. In this paper, we consider three different statistical approaches to tackle this problem: two of them, namely the well-known maximum likelihood classification (ML) and the support vector machine (SVM), are noncontextual methods. The third one, iterated conditional modes (ICM), exploits spatial context by using a Markov random field. We apply these methods to Landsat 5 Thematic Mapper (TM) data from Tenerife, the largest of the Canary Islands. Due to the size and the strong relief of the island, ground truth data could be collected only sparsely by examination of test areas for previously defined land cover classes.We show that after application of an unsupervised clustering method to identify subclasses, all classification algorithms give satisfactory results (with statistical overall accuracy of about 90%) if the model parameters are selected appropriately. Although being superior to ML theoretically, both SVM and ICM have to be used carefully: ICM is able to improve ML, but when applied for too many iterations, spatially small sample areas are smoothed away, leading to statistically slightly worse classification results. SVM yields better statistical results than ML, but when investigated visually, the classification result is not completely satisfying. This is due to the fact that no a priori information on the frequency of occurrence of a class was used in this context, which helps ML to limit the unlikely classes.     

A fuzzy AHP-DEA approach for multiple criteria ABC inventory classification

In order to efficiently control the inventory items and determine the suitable ordering policies for them, multi-criteria ABC inventory classification, which is one of the most common techniques of production and inventory control, is used. In this classification, other criteria in addition to annual dollar usage are taken into account and then the items are classified in three classes with different ordering policies, based on their priority. In this paper, we propose an integrated fuzzy analytic hierarchy process-data envelopment analysis (FAHP-DEA) for multiple criteria ABC inventory classification. The proposed FAHP–DEA methodology uses the FAHP to determine the weights of criteria, linguistic terms such as Very High, High, Medium, Low and Very Low to assess each item under each criterion, the data envelopment analysis (DEA) method to determine the values of the linguistic terms, and the simple additive weighting (SAW) method to aggregate item scores under different criteria into an overall score for each item. The integrated FAHP–DEA methodology is illustrated using a real case study.     

Extraction of training samples from time-series MODIS imagery and its utility for land cover classification

A number of classification techniques to generate land cover maps from satellite imagery have been proposed but supervised classification with manual selection and delineation of training samples (TSs) continues to be the preferred technique. The current practices of field visits and manual delineation of TSs by visual recognition are highly demanding on both resources and time, with limited utility. With an increase in the number of Earth Observation Satellite (EOS) platforms and the enormous data that they generate, there is a need to process the data quickly and efficiently for creating global science products. Towards this goal, an attempt has been made in this article to develop a method for the automatic extraction of the TSs from the time series of Moderate Resolution Imaging Spectroradiometer (MODIS – 250 m) vegetation index (VI), which can then be used for supervised classification to create a land cover map with any classification technique on relevant remotely sensed data. The TSs contained 1.27%, 0.09% and 1.18% of the total pixels for the forest, crop and water classes of the study region. Validation with Advanced Wide Field Sensor (AWiFS – 56 m)-derived national land use/land cover (LULC) map of India shows a complete agreement with the location of what can be considered as pure class pixels. The article also demonstrates and compares the utility of these TSs with an expert choice of TSs on MODIS time-series data using k-nearest neighbour, and support vector machine (SVM) classifiers and on a single-scene Linear Imaging Self-Scanning Sensor-3 (LISS-3 – 24 m) imagery using maximum likelihood (ML) classifier.     

Geostatistical and texture analysis of airborne-acquired images used in forest classification

Airborne sensor image texture derived following a geostatistical analysis can increase the accuracy of forest classification because the resulting texture is insensitive to random variations in spectral response but related to the structural features of interest at the scale of a forest inventory (e.g. tree species). The combination of spectral and textural data derived from a kriging surface provided 86% classification accuracy in 36 pure and mixed-wood stands in seven forest classes in Alberta. This is an increase over the classification accuracy obtained when texture was derived from the original image data, and when the spectral response patterns were used alone.     

Wetland change mapping for the U.S. mid-Atlantic region using an outlier detection technique

Although the impacts of wetland loss are often felt at regional scales, effective planning and management require a comparative assessment of local needs, costs, and benefits. Satellite remote sensing can provide spatially explicit, synoptic land cover change information to support such an assessment. However, a common challenge in conventional remote sensing change detection is the difficulty of obtaining phenologically and radiometrically comparable data from the start and end of the time period of interest. An alternative approach is to use a prior land cover classification as a surrogate for historic satellite data and to examine the self-consistency of class spectral reflectances in recent imagery. We produced a 30-meter resolution wetland change probability map for the U.S. mid-Atlantic region by applying an outlier detection technique to a base classification provided by the National Wetlands Inventory (NWI). Outlier-resistant measures – the median and median absolute deviation – were used to represent spectral reflectance characteristics of wetland class populations, and formed the basis for the calculation of a pixel change likelihood index. The individual scene index values were merged into a consistent region-wide map and converted to pixel change probability using a logistic regression calibrated through interpretation of historic and recent aerial photography. The accuracy of a regional change/no-change map produced from the change probabilities was estimated at 89.6%, with a Kappa of 0.779. The change probabilities identify areas for closer inspection of change cause, impact, and mitigation potential. With additional work to resolve confusion resulting from natural spatial heterogeneity and variations in land use, automated updating of NWI maps and estimates of areal rates of wetland change may be possible. We also discuss extensions of the technique to address specific applications such as monitoring marsh degradation due to sea level rise and mapping of invasive species.     

Improvements in land use mapping for irrigated agriculture from satellite sensor data using a multi-stage maximum likelihood classification

The accuracy of conventional land use classification of irrigated agriculture from optical satellite images using maximum likelihood supervised classification was compared with a classification based on multistage maximum likelihood supervised classification. In the multistage maximum likelihood classification series of sub-classifications were carried out which included masking and/or omitting certain crops from the classifications. These series of classifications improved the identification of individual crops/land use types. The output from the optimum sub-classifications were stacked to give an overall crop types/land use map. When the multistage classification was tested against a single stage classification on a large irrigation scheme in Central Asia the final accuracy of crop/land use classification increased from 85% to 94%. Field verification confirmed the accuracy at 93.5%. These results were achieved with a single Landsat 7 Enhanced Thematic Mapper (ETM+) sensor dataset as of 2 August 1999 over an area of 38.5?km².