The mixed pixel problem is common in remote sensing. A soft classification can generate land cover class fraction images that illustrate the areal proportions of the various land cover classes within pixels. The spatial distribution of land cover classes within each mixed pixel is, however, not represented. Super-resolution land cover mapping (SRM) is a technique to predict the spatial distribution of land cover classes within the mixed pixel using fraction images as input. Spatial–temporal SRM (STSRM) extends the basic SRM to include a temporal dimension by using a finer-spatial resolution land cover map that pre- or postdates the image acquisition time as ancillary data. Traditional STSRM methods often use one land cover map as the constraint, but neglect the majority of available land cover maps acquired at different dates and of the same scene in reconstructing a full state trajectory of land cover changes when applying STSRM to time-series data. In addition, the STSRM methods define the temporal dependence globally, and neglect the spatial variation of land cover temporal dependence intensity within images. A novel local STSRM (LSTSRM) is proposed in this paper. LSTSRM incorporates more than one available land cover map to constrain the solution, and develops a local temporal dependence model, in which the temporal dependence intensity may vary spatially. The results show that LSTSRM can eliminate speckle-like artifacts and reconstruct the spatial patterns of land cover patches in the resulting maps, and increase the overall accuracy compared with other STSRM methods.
Conventional image classification routines are often inappropriate for the mapping of continuous phenomena suchy as heathland vegetation. To allow for the natural fuzziness of such an environment, a fuzzy sets algorithm may be used to model the heathland vegetation more appropriately than a classification. The results of this study show that the fuzzy c-means algorithm can be used to discriminate accurately between the end points of a set of continua, and that class membership functions derived from the analysis are sensitive to the botanical composition of the vegetation canopy. Mapping the fuzzy membership functions will therefore enable a more realistic portrayal of the healthland vegetation than a conventional classification.
Google Earth Engine (GEE) provides a convenient platform for applications based on optical satellite imagery of large areas. With such data sets, the detection of cloud is often a necessary prerequisite step. Recently, deep learning-based cloud detection methods have shown their potential for cloud detection but they can only be applied locally, leading to inefficient data downloading time and storage problems. This letter proposes a method to directly perform cloud detection in Landsat-8 imagery in GEE based on deep learning (DeepGEE-CD). A deep neural network (DNN) was first trained locally, and then the trained DNN was deployed in the JavaScript client of GEE. An experiment was undertaken to validate the proposed method with a set of Landsat-8 images and the results show that DeepGEE-CD outperformed the widely used function of mask (Fmask) algorithm. The proposed DeepGEE-CD approach can accurately detect cloud in Landsat-8 imagery without downloading it, making it a promising method for routine cloud detection of Landsat-8 imagery in GEE.
More than half of all tropical forests are degraded by human impacts, leaving them threatened with conversion to agricultural plantations and risking substantial biodiversity and carbon losses. Restoration could accelerate recovery of aboveground carbon density (ACD), but adoption of restoration is constrained by cost and uncertainties over effectiveness. We report a long-term comparison of ACD recovery rates between naturally regenerating and actively restored logged tropical forests. Restoration enhanced decadal ACD recovery by more than 50%, from 2.9 to 4.4 megagrams per hectare per year. This magnitude of response, coupled with modal values of restoration costs globally, would require higher carbon prices to justify investment in restoration. However, carbon prices required to fulfill the 2016 Paris climate agreement [$40 to $80 (USD) per tonne carbon dioxide equivalent] would provide an economic justification for tropical forest restoration.
No abstract is provided for this article.
In some classifications the importance of classes varies and it is desirable to weight allocation to selected classes. This is common in classifications of remotely sensed imagery, especially as class occurrence can vary markedly. If, for instance, there is prior knowledge on the distribution of class occurrence this weighting can be achieved with widely used statistical classifiers by setting appropriate a priori probabilities of class membership. With an arificial neural network the incorporation of prior knowledge is more problematic. An approach to weight class allocation in an artificial neural network classifcation by replicating selected training patterns is discussed. In comparison against a discriminant analysis for the classification of synthetic aperture radar imagery the results showed that training pattern replication could be used to weight class allocation with an effect similar to that of incorporating a priori probabilities of class membership into the discriminant analysis and resulted in a significant increase in classification accuracy.
In this paper the authors address the problem of interpreting and classifying aggregate data sources and draw parallels between tasks commonly encountered in image processing and census analysis. Both of these fields already have a range of standard classification tools which are applied in such situations, but these are hindered by the aggregate nature of the input data. An approach to ‘unmixing’ aggregate data, and thus to revealing the nature of the subunit variation masked by aggregation, is introduced. This approach has already shown considerable success in Earth Observation applications, and in this paper the authors present the adaptation and application of the approach to Census small area statistics data for Southampton, Hants, revealing something of the social composition of Southampton's enumeration districts. The unmixing technique utilises an artificial neural network.
Forest biophysical properties are typically estimated and mapped from remotely sensed data through the application of a vegetation index. This generally do
Thematic maps are often derived from remotely sensed imagery via a supervised image classification analysis. The training and testing stages of a supervised image classification may proceed ignorant of the presence of some classes in the region to be mapped. This violates the assumption of an exhaustively defined set of classes that is often made in classification analyses. In such circumstances, the overall accuracy of a thematic map produced by the application of a trained classifier will be less than the accuracy of the classification of the test set by the same classifier. This situation arises because the cases of an untrained class can normally only be commissioned into the set of trained classes. Simple mathematical relationships between classification and map accuracy are shown for assessments of overall, user's and producer's accuracy. For example, it is shown that in a simple scenario the accuracy of a thematic map is less than that of a classification, scaling as a function of the abundance of the untrained class(es). Impacts on other estimates made from thematic maps, such as class areal extent, are also briefly discussed. When using a thematic map, care is needed in interpreting and using classification accuracy assessments as sometimes they may not reflect properties of the map well.
Quantifying the above ground biomass of tropical forests is critical for understanding the dynamics of carbon fluxes between terrestrial ecosystems and the atmosphere, as well as monitoring ecosystem responses to environmental change. Remote sensing remains an attractive tool for estimating tropical forest biomass but relationships and methods used at one site have not always proved applicable to other locations. This lack of a widely applicable general relationship limits the operational use of remote sensing as a method for biomass estimation, particularly in high biomass ecosystems. Here, multispectral Landsat TM and JERS-1 SAR data were used together to estimate tropical forest biomass at three separate geographical locations: Brazil, Malaysia and Thailand. Texture measures were derived from the JERS-1 SAR data using both wavelet analysis and Grey Level Co-occurrence Matrix methods, and coupled with multispectral data to provide inputs to artificial neural networks that were trained under four different training scenarios and validated using biomass measured from 144 field plots. When trained and tested with data collected from the same location, the addition of SAR texture to multispectral data showed strong correlations with above ground biomass (r =0.79, 0.79 and 0.84 for Thailand, Malaysia and Brazil respectively). Also, when networks were trained and tested with data from all three sites, the strength of correlation (r =0.55) was stronger than previously reported results from the same sites that used multispectral data only. Uncertainty in estimating AGB from different allometric equations was also tested but found to have little effect on the strength of the relationships observed. The results suggest that the inclusion of SAR texture with multispectral data can go someway towards providing relationships that are transferable across time and space, but that further work is required if satellite remote sensing is to provide robust and reliable methodologies for initiatives such as Reducing Emissions from Deforestation and Degradation (REDD+).
Conventional approaches to training a supervised image classification aim to fully describe all of the classes spectrally. To achieve a complete description of each class in feature space, a large training set is typically required. It is not, however, always necessary to have training statistics that provide a complete and representative description of the classes, especially if using nonparametric classifiers. For classification by a support vector machine, only the training samples that are support vectors, which lie on part of the edge of the class distribution in feature space, are required; all other training samples provide no contribution to the classification analysis. If regions likely to furnish support vectors can be identified in advance of the classification, it may be possible to intelligently select useful training samples. The ability to target useful training samples may allow accurate classification from small training sets. This potential for intelligent training sample collection was explored for the classification of agricultural crops from multispectral satellite sensor data. With a conventional approach to training, only a quarter of the training samples acquired actually made a positive contribution to the analysis and allowed the crops to be classified to a high accuracy (92.5%). The majority of the training set, therefore, was unnecessary as it made no contribution to the analysis. Using ancillary information on soil type, however, it would be possible to constrain the training sample acquisition process. By limiting training sample acquisition only to regions with a specific soil type, it was possible to use a small training set to classify the data without loss of accuracy. Thus, a small number of intelligently selected training samples may be used to classify a data set as accurately as a larger training set derived in a conventional manner. The results illustrate the potential to direct training data acquisition strategies to target the most useful training samples to allow efficient and accurate image classification.
The absence of assumptions about the dataset to be classified is one of the major attractions of neural networks for supervised image classification applications. Classification by a neural network does, however, make assumptions about the classes. One key assumption typically made is 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. This was observed in land cover classifications derived with multi-layer perceptron (MLP) and radial basis function (RBF) neural networks in which the presence of an untrained class resulted in a ∼12.5% decrease in the accuracy of crop classifications derived from airborne thematic mapper data. However, since the RBF network partitions feature space locally rather than globally as with the MLP, it was possible to reduce the commission of atypical cases into the set of trained classes through the setting of post-classification thresholds on the RBF network's outputs. As a result it was possible to identify and exclude some cases of untrained classes from a classification with a RBF network which resulted in an increase in classification accuracy.
The accuracy of an image classification is a function of a range of variables. To select an appropriate classification approach for a particular set of data the analyst must be aware of the significant variables which may affect classification accuracy and the nature of their effect. The effect of a variable, or small number of variables, on classification accuracy may be evaluated by straightforward comparison of classification accuracies. However, for the evaluation of the simultaneous effect of a large number of variables such an approach may be impractical. In such circumstances log-linear modelling may be used to identify the significant variables affecting classification accuracy and the nature of the effect of the significant variables elucidated from further analysis. Log-linear modelling was used here to evaluate the effect of four variables (training set size, waveband combination, classification algorithm and testing set size) on classification accuracy. Since the analyst usually has most control over the choice of classification algorithm most attention was focused on the effects of the other three variables on the accuracies of classifications derived from conventional probabilistic, fuzzy set and neural network classification algorithms. The results showed that these classification algorithms were sensitive to variations in the other three variables. Overall the neural network classifications were generally the most accurate. The accuracies of the neural network classifications were, however, most influenced by training set size, with higher accuracies obtained with the use of large training sets. Alternatively, the other classification algorithms were least affected by the training set size and more sensitive to the testing set size and the waveband combination used. The results should help an analyst design an appropriate approach for an image classification.
This paper presents an approach to the classification of crop type using multi‐temporal airborne synthetic aperture radar (SAR) data. Following radiometric correction of the data the accuracy of a per‐field crop classification reached 90% for three classes using data acquired on four dates. A comparable accuracy of 88% could be obtained for a classification of the same classes using data acquired on only two dates. Increasing the number of classes from three to seven reduced the classification accuracies to 55% and 69% when using data from two and four dates respectively.
Combining super-resolution techniques can increase the accuracy with which the shape of objects may be characterised from imagery. This is illustrated with two approaches to combining the contouring and pixel swapping methods of super-resolution mapping for binary classification applications. In both approaches, the output of the pixel swapping method is softened to allow a contour of equal class membership to be fitted to it to represent the inter-class boundary. The accuracy of super-resolution mapping with the individual and combined techniques is explored, including an assessment of the effect of variation in the number of neighbors and zoom factor on pixel swapping based analyses. When combined, the error with which objects of varying shape were represented was typically greatly reduced relative to that observed from the application of the methods individually. For example, the root mean square error in mapping the boundary of an aeroplane represented in relatively fine spatial resolution imagery decreased from 14.41 m with contouring and 4.35 m with pixel swapping to 3.07 m when the approaches were combined.
