The design of the training stage of a supervised classification should account for the properties of the classifier to be used. Consideration of the way the classifier operates may enable the training stage to be designed in a manner which ensures that the aim of the classification is satisfied with the use of a small, inexpensive, training set. It may, therefore, be possible to reduce the training set size requirements from that generally expected with the use of standard heuristics. Substantial reductions in training set size may be possible if interest is focused on a single class. This is illustrated for mapping cotton in north-western India by support vector machine type classifiers. Four approaches to reducing training set size were used: intelligent selection of the most informative training samples, selective class exclusion, acceptance of imprecise descriptions for spectrally distinct classes and the adoption of a one-class classifier. All four approaches were able to reduce the training set size required considerably below that suggested by conventional widely used heuristics without significant impact on the accuracy with which the class of interest was classified. For example, reductions in training set size of ∼90% from that suggested by a conventional heuristic are reported with the accuracy of cotton classification remaining nearly constant at ∼95% and ∼97% from the user's and producer's perspectives respectively.
Small water bodies (SWBs), such as ponds and on-farm reservoirs, are a key part of the hydrological system and play important roles in diverse domains from agriculture to conservation. The monitoring of SWBs has been greatly facilitated by medium-spatial-resolution satellite images, but the monitoring accuracy is considerably affected by the mixed-pixel problem. Although various spectral unmixing methods have been applied to map sub-pixel surface water fractions for large water bodies, such as lakes and reservoirs, it is challenging to map SWBs that are small in size relative to the image pixel and have dissimilar spectral properties. In this study, a novel regression-based surface water fraction mapping method (RSWFM) using a random forest and a synthetic spectral library is proposed for mapping 10 m spatial resolution surface water fractions from Sentinel-2 imagery. The RSWFM inputs a few endmembers of water, vegetation, impervious surfaces, and soil to simulate a spectral library, and considers spectral variations in endmembers for different SWBs. Additionally, RSWFM applies noise-based data augmentation on pure endmembers to overcome the limitation often arising from the use of a small set of pure spectra in training the regression model. RSWFM was assessed in ten study sites and compared with the fully constrained least squares (FCLS) linear spectral mixture analysis, multiple endmember spectral mixture analysis (MESMA), and the nonlinear random forest (RF) regression without data-augmentation. The results showed that RSWFM decreases the water fraction mapping errors by ~ 30%, ~15%, and ~ 11% in root mean square error compared with the linear FCLS, MESMA unmixings, and the nonlinear RF regression without data-augmentation respectively. RSWFM has an accuracy of approximately 0.85 in R2 in estimating the area of SWBs smaller than 1 ha.
The full realization of the potential of remote sensing as a source of environmental information requires an ability to generalize in space and time. Here, the ability to generalize in space was investigated through an analysis of the transferability of predictive relations for the estimation of tropical forest biomass from Landsat TM data between sites in Brazil, Malaysia and Thailand. The data sets for each test site were acquired and processed in a similar fashion to facilitate the analyses. Three types of predictive relation, based on vegetation indices, multiple regression and feedforward neural networks, were developed for biomass estimation at each site. For each site, the strongest relationships between the biomass predicted and that measured from field survey was obtained with a neural network developed specifically for the site (r>0.71, significant at the 99% level of confidence). However, with each type of approach problems in transferring a relation to another site were observed. In particular, it was apparent that the accuracy of prediction, as indicated by the correlation coefficient between predicted and measured biomass, declined when a relation was transferred to a site other than that upon which it was developed. Part of this problem lies with the observed variation in the relative contribution of the different spectral wavebands to predictive relations for biomass estimation between sites. It was, for example, apparent that the spectral composition of the vegetation indices most strongly related to biomass differed greatly between the sites. Consequently, the relationship between predicted and measured biomass derived from vegetation indices differed markedly in both strength and direction between sites. Although the incorporation of test site location information into an analysis resulted in an increase in the strength of the relationship between predicted and actual biomass, considerable further research is required on the problems associated with transferring predictive relations.
Ground-data quality is an important issue in the analysis of remotely sensed data. Since the results of such an analysis are evaluated against the ground data these ground data must be accurate. Two issues affecting the quality of soil moisture content ground data are investigated in this paper. These are the laboratory procedures employed in its estimation and the effect of cultivation practices on the distribution of surface soil moisture in the field. Laboratory procedures commonly used may not always be appropriate and different approaches are likely to give estimates of variable accuracy. Cultivation practices were found to produce significant variations in soil moisture distribution, with over 10 percent difference in soil moisture content estimates between the top and bottom of a ridged soil. Ground data collectors using different sampling designs and adopting dissimilar laboratory techniques may therefore produce considerably different estimates of the soil moisture content. Consequently ground-data collection programmes should be fully documented, especially if the results obtained by different research groups are to be compared or merged, so that users may be aware of the quality of the data recorded
Gathering information from remotely sensed data for input to geographical information systems is a relatively common practice. In situations where the remotely sensed data are used indirectly and it is a thematic classification of the data which is incorporated into the system care is required in assessing its quality. Evaluating the classification in relation to a pre‐defined threshold accuracy can be unreliable since classification accuracy is a spatially variable phenomenon. In constructing a GIS it could be possible to include erroneously an unsuitable classification. Conversely a suitable classification could be excluded. Users of thematic classifications derived from remotely sensed data therefore need to be aware of the magnitude and spatial distribution of classification accuracy for its optimal use in the system.
No abstract is provided for this article.
Synthetic aperture radar "SAR" systems are an attractive source of information for agricultural crop classification applications, particularly in regions where cloud cover is a problem. The accuracy with which crops can be classified is dependent on a range of sensor properties, including the SAR operating configuration. This paper focuses on the effect of one aspect of the SAR operating configuration, polarization, on crop classification accuracy using uncalibrated C-band polarimetric SAR data. Conventional like- and cro spolarized configurations and polarimetric coefficients "the pedestal and variation coefficients" were used as discriminating variables in classifications of agricultural crops. Two approaches to classification were investigated, a discriminant analysis and an artif cial neural network and results from a set of training classifications are presented. The results show that the polarimetric coefficients used provided a high level of inter-class discrimination and that a nine-class classification with an accuracy of up to 78·75 per cent could be produced from these C-band polarimetric SAR data. Classification accuracy was also influenced by the classification technique, with the neural network achieving significantly higher levels of inter-class separability in the training data than the discriminant analysis.
The relationship between middle infrared reflectance and various scenarios of preceeding rainfall for a range of tropical forest types is investigated. Statistically significant correlations for the relationship were generally observed when the rainfall data were acquired over a period of about a month with a short time lag before image acquisition.
Accuracy assessment should be a fundamental part of a programme that maps land cover from remotely sensed imagery but this activity is often constrained by the lack of high quality ground reference data. Here, two sources of volunteered data are used to illustrate the potential of amateur or neogeographical activity in map validation. Ground based photographs acquired from an internet-based collaborative project and interpreted by a set of four further volunteers provided the reference data to support evaluation of the Globcover map's representation of tropical forest in West Africa. Although the results highlight some concerns with volunteered data, notably the low levels of inter-volunteer agreement they also show that such imperfect data may be used to derive credible estimates on accuracy from both a site and non-site specific perspective. Specifically, the estimates of the producer's accuracy of forest and of forest extent derived from the free and volunteered data using a latent class model were of comparable magnitude to those derived in a formal validation by experts; the estimate of forest extent was within 1.38-9.08% of reference estimates while the difference in estimated producer's accuracy from that derived in an authoritative assessment of map accuracy was 2.82% and 0.34% for the forest and non-forest classes respectively.
No abstract is provided for this article.
While this final chapter concludes this book, the subject of this chapter (GeoDynamics) is in its infancy. Thus, this chapter provides a general review of some of the most important research strands emerging within GeoDynamics (as evident from the chapters of this book) and a look forward to the future. The objective of the chapter is to encourage researchers to explore and model Earth surface processes and predict their outcomes using the tools of GeoDynamics.
Freedom from assumptions about the data set used is one attraction of neural network classifiers. However, neural network classification is not assumption-free. It is typically assumed that the set of classes has been defined exhaustively. If this assumption is unsatisfied, cases of an untrained class will be present and commissioned into the set of trained classes to the detriment of classification accuracy, for both hard and soft classifications. This is illustrated with MLP and RBF neural networks together with suggestions of how to reduce the problem for both hard and soft image classifications.
Radiometric balancing is a form of radiometric correction that can be applied to radar imagery. However, its application can be based on untenable assumptions about the characteristics of the scene and so can be unreliable. The method of radiometrically balancing an image and the requirements of the technique are outlined along with an example of where its use would be inappropriate. A recommendation is made against policies of pre-dissemination radiometric balancing and it is suggested that users should evaluate the use of other techniques which may be more suitable for their imagery.
Prior knowledge of class occurrence can be used in conventional statistical classifications to increase classification accuracy significantly. These classifications, however, require typically a large training set for their correct use. When small training sets are available therefore an alternative classification technique is required and one increasingly popular alternative is the use of artificial neural network techniques. Whilst many comparative studies have shown artificial neural network classifications to be more accurate than those derived from conventional statistical classifiers the difference in accuracy can be more than made-up if prior knowledge was incorporated in the statistical classification. To make the full use of artificial neural networks therefore approaches for the incorporation of prior knowledge must be developed. This paper proposes one approach which may be used and illustrates its use in the classification of agricultural crops from synthetic aperture radar data with a minimal training set. The results show that whilst the artificial neural network could accurately learn the training data the classification of an independent testing set was poor, an accuracy of only 27.0 per cent was derived for a seven class classification. Incorporating prior knowledge, however, significantly increased classification accuracy to 58.4 per cent. This latter result was comparable, in terms of accuracy and the pattern of class allocation, to a classification of the same data set by a discrimination analysis with prior knowledge.
Conventional supervised classifiers cannot accommodate mixed pixels directly but may be modified to do so throughout the classification process. Here mixed pixels are included in all three stages of maximum likelihood and neural network classifications. The results show that by accommodating for mixed pixels in the classification, more accurate, appropriate and useful outputs may be derived.
Methods for mapping the shoreline at a sub-pixel level are evaluated. The most accurate predictions of shoreline location were made from an approach based on simulated annealing applied to the output of a soft classification (RMSE=2.25 m).
