Sucharita Gopal

Global land cover mapping from MODIS: algorithms and early results

Until recently, advanced very high-resolution radiometer (AVHRR) observations were the only viable source of data for global land cover mapping. While many useful insights have been gained from analyses based on AVHRR data, the availability of moderate resolution imaging spectroradiometer (MODIS) data with greatly improved spectral, spatial, geometric, and radiometric attributes provides significant new opportunities and challenges for remote sensing-based land cover mapping research. In this paper, we describe the algorithms and databases being used to produce the MODIS global land cover product. This product provides maps of global land cover at 1-km spatial resolution using several classification systems, principally that of the IGBP. To generate these maps, a supervised classification methodology is used that exploits a global database of training sites interpreted from high-resolution imagery in association with ancillary data. In addition to the IGBP class at each pixel, the MODIS land cover product provides several other parameters including estimates for the classification confidence associated with the IGBP label, a prediction for the most likely alternative class, and class labels for several other classification schemes that are used by the global modeling community. Initial results based on 5 months of MODIS data are encouraging. At global scales, the distribution of vegetation and land cover types is qualitatively realistic. At regional scales, comparisons among heritage AVHRR products, Landsat TM data, and results from MODIS show that the algorithm is performing well. As a longer time series of data is added to the processing stream and the representation of global land cover in the site database is refined, the quality of the MODIS land cover product will improve accordingly.

FUZZY SET THEORY AND THEMATIC MAPS: ACCURACY ASSESSMENT AND AREA ESTIMATION

Traditionally, the classes in thematic maps have been treated as crisp sets, using classical set theory. In this formulation, map classes are assumed to be mutually exclusive and exhaustive. This approach limits the ability of thematic maps to represent the continuum of variation found in most landscapes. Substitution of fuzzy sets allows more flexibility for treatment of map classes in the areas of accuracy assessment and area estimation. Accuracy assessment methods based on fuzzy sets allow consideration of the magnitude of errors and assessment of the frequency of ambiguity in map classes. An example of an accuracy assessment from a vegetation map of the Plumas National Forest illustrates the implementation of these methods. Area estimation based on fuzzy sets and using accuracy assessment data allows estimation of the area of classes as a function of levels of class membership. The fuzzy area estimation methods are an extension of previous methods presented by Card (1982). One interesting result is that the sum of the areas of the classes in a map need not be unity. This approach allows a wider range of queries within a GIS.

Remote sensing of forest change using artificial neural networks

A prolonged drought in the Lake Tahoe Basin in California has resulted in extensive conifer mortality. This phenomenon can be analyzed using (multitemporal) remote sensing data. Prior research in the same region used more traditional methods of change detection. The present paper introduces a third approach to change detection in remote sensing based on artificial neural networks. The neural network architecture used is a multilayer feedforward network. The results of the study indicate that the artificial neural network (ANN) estimates conifer mortality more accurately than the other approaches. Further, an analysis of its architecture reveals that it uses identifiable scene characteristics-the same as those used by a Gramm-Schmidt transformation. ANN models offer a viable alternative for change detection in remote sensing.

ART neural networks for remote sensing: vegetation classification from Landsat TM and terrain data

A new methodology for automatic mapping from Landsat thematic mapper (TM) and terrain data, based on the fuzzy ARTMAP neural network, is developed. System capabilities are tested on a challenging remote sensing classification problem, using spectral and terrain features for vegetation classification in the Cleveland National Forest. After training at the pixel level, system performance is tested at the stand level, using sites not seen during training. Results are compared to those of maximum likelihood classifiers, as well as back propagation neural networks and K nearest neighbor algorithms. ARTMAP dynamics are fast, stable, and scalable, overcoming common limitations of back propagation. Best results are obtained using a hybrid system based on a convex combination of fuzzy ARTMAP and maximum likelihood predictions. A prototype remote sensing example introduces each aspect of data processing and fuzzy ARTMAP classification. The example shows how the network automatically constructs a minimal number of recognition categories to meet accuracy criteria. A voting strategy improves prediction and assigns confidence estimates by training the system several times on different orderings of an input set.

Mapping forest vegetation using landsat TM imagery and a canopy reflectance model

Abstract Estimates of mean tree size and cover for each forest stand from an invertible forest canopy reflectance model are part of a new forest vegetation mapping system. Image segmentation defines stands which are sorted into general growth forms using per-pixel image classifications. Ecological models based on terrain relations predict species associations for the conifer, hardwood, and brush growth forms. The combination of the model-based estimates of tree size and cover with species associations yields general-purpose vegetation maps useful for a variety of land management needs. Results of timber inventories in the Tahoe and Stanislaus National Forests indicate the vegetation maps form a useful basis for stratification. Patterns in timber volumes for the strata reveal that the cover estimates are more reliable than the tree size estimates. A map accuracy assessment of the Stanislaus National Forest shows high overall map accuracy and also illustrates the problems in estimating tree size.

A Neural Network Method for Mixture Estimation for Vegetation Mapping

While most forest maps identify only the dominant vegetation class in delineated stands, individual stands are often better characterized by a mix of vegetation types. Many land management applications, including wildlife habitat studies, can benefit from knowledge of mixes. This article examines various algorithms that use data from the Landsat Thematic Mapper (TM) satellite to estimate mixtures of vegetation types within forest stands. Included in the study are maximum likelihood classification and linear mixture models as well as a new methodology based on the ARTMAP neural network. Two paradigms are considered: classification methods, which describe stand-level vegetation mixtures as mosaics of pixels, each identified with its primary vegetation class; and mixture methods, which treat samples as blends of vegetation, even at the pixel level. Comparative analysis of these mixture estimation methods, tested on data from the Plumas National Forest, yields the following conclusions: 1) Accurate estimates of proportions of hardwood and conifer cover within stands can be obtained, particularly when brush is not present in the understory; 2) ARTMAP outperforms statistical methods and linear mixture models in both the classification and the mixture paradigms; 3) topographic correction fails to improve mapping accuracy; and 4) the new ARTMAP mixture system produces the most accurate overall results. The Plumas data set has been made available to other researchers for further development of new mapping methods and comparison with the quantitative studies presented here, which establish initial benchmark standards.

Fuzzy neural network classification of global land cover from a 1° AVHRR data set

Abstract Phenological differences among broadly defined vegetation types can be a basis for global scale landcover classification at a very coarse spatial scale. Using an annual sequence of composited normalized difference vegetation index (NDVI) values from AVHRR data set composited to 1° DeFries and Townshend (1994) classified eleven global land-cover types with a maximum likelihood classifier. Classification of these same data using a neural network architecture called fuzzy ARTMAP indicate the following: i) When fuzzy ARTMAP is trained using 80% of the data and tested on the remaining (unseen) 20% of the data, classification accuracy is more than 85% compared with 78% using the maximum likelihood classifier; ii) classification accuracies for various splits of training/testing data show that an increase in the size of training data does not result in improved accuracies; iii) classification results vary depending on the use of latitude as an input variable similar to the results of DeFries and Townshend; and iv) fuzzy ARTMAP dynamics including a voting procedure and the number of internal nodes can be used to describe uncertainty in classification. This study shows that artificial neural networks are a viable alternative for global scale landcover classification due to increased accuracy and the ability to provide additional information on uncertainty.

Navigator: A psychologically based model of environmental learning through navigation

Abstract This paper describes an implementation of a computational process model of spatial learning. The model is designed to represent basic components of human information processing, as identified by contemporary psychological research and theory. The model comprises two modules, representing an objective view of a suburban environment and a cognitive view of the environment, together with the associated cognitive processes relating to spatial learning and navigation. The second module is based upon information-processing assumptions that relate, for example, to multiple sites of storage, filtering and selection, sensitivity to the importance of environmental features, and forgetting. The implications of the model concerning the processing stages that limit spatial learning, the incremental nature of learning, and the effects of both individual and environmental variation are tested in a series of simulations. By this means the model is of value in tracing out how the architecture of human information processing, in a broad sense, might limit and control the extraction and use of environmental information. More specific model-based predictions are also considered.

Forest mapping with a generalized classifier and Landsat TM data

Monitoring landcover and landcover change at regional and global scales often requires Landsat data to identify and map landscape features and patterns with sufficient detail. Analytical methods based on image-by-image interpretation are too time-consuming and labor-intensive for studies of large areas to be undertaken with any degree of frequency. One potential solution is to develop algorithms or classifiers that can be generalized beyond the arena of the initial training to new images from different spatial, temporal or sensor domains. Building upon earlier success with a generalized classifier to monitor forest change, we now address the question of generalization for classifications of stable landcovers. We evaluate the ability of a supervised neural network, Fuzzy ARTMAP, to identify conifer forest across time and space with Landsat Thematic Mapper (TM) images for a region in northwest Oregon. We also assess the effects of atmospheric corrections on generalized classification accuracies. Using midsummer images atmospherically corrected with a simple dark-object-subtraction (DOS) method, there is no statistically significant loss of accuracy as the classification is extended from the initial training image to other images from the same scene (path and row): temporal generalization is successful. Extending the classifier across space and time to nearby scenes results in a mean decline of 8–13% accuracy depending on the atmospheric correction used. Obvious sources of error, such as seasonality, solar angle variation, and complexity of landcover identification, do not explain the decline in error. Additionally, the patterns in generalization accuracies are complex, and the relationship between pairs of training and testing images is not necessarily reciprocal, i.e., good training data are not necessarily good testing data. Simple DOS atmospheric corrections produce classifications with comparable accuracies as classifications from the more complex radiative transfer corrections. These findings are based on over 200 classifications. A high degree of variability in the classification accuracies underscores the importance of extensive, in-depth analysis of remote sensing techniques and applications, and highlights the potential problem for misleading results based on just a few tests. Generalization is well suited for multitemporal classifications of one Landsat scene. Using simple DOS and midsummer images, generalization offers the opportunity for frequent landcover mapping of a Landsat scene without having to retrain the classifier for each time period of interest. However, at this point, the utility of regional landcover mapping with a generalized classifier remains limited.

Gaussian mixture discriminant analysis and sub-pixel land cover characterization in remote sensing

Mixture analysis is a necessary component for capturing sub-pixel heterogeneity in the characterization of land cover from remotely sensed images. Mixture analysis approaches in remote sensing vary from conventional linear mixture models to nonlinear neural network mixture models. Linear mixture models are fairly simple and generally result in poor mixture analysis accuracy. Neural network models can achieve much higher accuracy, but typically lack interpretability. In this paper we present a mixture discriminant analysis (MDA) model for inferring land cover fractions within forest stands from Landsat Thematic Mapper images. Specifically, individual class distributions are modeled as mixtures of subclasses of Gaussian distributions, and land cover fractions are estimated using the corresponding posterior probabilities. Compared to a benchmark study on accuracy of mixture models with Plumas National Forest data, this MDA model easily outperforms traditional linear mixture models and parallels the performance of the ARTMAP neural network mixture model. In other words, the MDA model is observed to successfully combine the performance characteristics of more complex neural network models (due to the nonlinear nature of its classification rules), with the ease of interpretation associated with linear mixture models (due to its relatively simple structure). MDA models therefore offer an attractive alternative for addressing the mixture modeling problem in remote sensing.