Claudio Conese

Use of probability entropy for the estimation and graphical representation of the accuracy of maximum likelihood classifications

Abstract A method is proposed for statistically evaluating the accuracy levels of maximum likelihood classifications and representing them graphically. Based on the concept that the heterogeneity of maximum likelihood membership probabilities can be taken as an indicator of the confidence for the classification, such a parameter is estimated for all pixels as relative probability entropy and represented in a separate channel. After a brief presentation of the statistical basis of the methodology, this is applied to a conventional and two modified maximum likelihood classifications in a case study using Landsat TM scenes. The results demonstrate the efficiency of the approach and, particularly, its usefulness for operational applications.

Integration of High and Low Resolution NDVI Data for Monitoring Vegetation in Mediterranean Environments

Abstract The integration of the useful features of high and low spatial and temporal resolution satellite data is a major issue in remote sensing studies. The current work presents the development and testing of a procedure based on classification and regression analysis techniques for generating an NDVI data set with the spatial resolution of Landsat TM images and the temporal resolution of NOAA AVHRR maximum-value composites. The procedure begins with a classification of the high resolution TM data which yields land use references. These are degraded to low spatial resolution in order to produce abundance images comparable with the AVHRR data. Linear regressions are then applied between the AVHRR NDVI data and the abundance images to estimate the profiles of the pure classes, which are then merged to the high spatial resolution classification outputs to generate an integrated data set. Experiments carried out in an area of Tuscany (Central Italy) intercomparing different strategies for each methodological step (hard and fuzzy classification, mean and Gaussian degradation, uni- and multivariate regression) identified an optimum methodology composed of fuzzy classification, mean degradation, and multivariate regression procedures.

Selection of optimum bands from TM scenes through mutual information analysis

Abstract The selection of optimum band subsets from remotely sensed for visual interpretation or automatic processing is an interesting task which will assume growing importance with the availability of highly multispectral data from future sensors. The usual methods for the mathematical evaluation of the best combination of channels are based on parametric statistical procedures such as eigenvector analysis and calculation of separability measurements. These procedures are not easy to be interpreted or computationally expensive and are not suited for evaluating the probabilistic information which can be exploited by non-parametric processes. For this kind of application, a method based on mutual information analysis is put forward in the present paper. Mutual information analysis is a statistical procedure which, using the concept of system entropy, is capable of mathematically evaluating the probabilistic information common to different variables. When applied to remotely sensed scenes superimposed on ground references related to some theme (for example vegetation types), information analysis can indicate which channels express more information about that theme. The method was applied to some Landsat TM scenes from three Italian areas about which ground references were available. Some mixed parametric-non-parametric classifications were then performed to test the results of the analyses. From these tests the subsets identified were demonstrated to be significantly more informative than standard subsets, which testifies to the efficiency of the procedure.

Use of error matrices to improve area estimates with maximum likelihood classification procedures

Abstract The maximum likelihood classifier is by far the most widespread among supervised classification methods. This procedure offers numerous advantages, but it has considerable shortcomings in the presence of strongly irregular spectral distributions, mainly related to bias in area estimates. Since these cases are quite common, some methods have already been proposed to correct biased area estimates from maximum likelihood classifications, but they are often not generally applicable or statistically stable. In this article a method is put forward to correct maximum likelihood assignment probabilities by means of a transition matrix; this matrix is derived through a simple mathematical algorithm from a contingency table of a previous classification compared to reference pixels. The purpose is clearly to attain a diagonalization of the final error sources to better estimate area extents and, above all, to achieve higher global discrimination accuracy. As different environmental situations may cause wide variability in the results of such a procedure, this was tested in three case studies using TM data acquired over areas with different landscapes. The results, evaluated by means of suitable statistics, significantly support that the method has general validity and applicability.

Estimation of forest parameters through fuzzy classification of TM data

Several studies have investigated the utility of Landsat 5 TM imagery to estimate forest parameters such as stand composition and density. Regression equations have generally been used to relate these parameters to the radiance responses of the TM channels. Such a method is not feasible in highly complex landscapes, where forest mixtures and terrain irregularities may obscure the existence of simple relationships. A fuzzy approach to the problem is presented based on a multi-step procedure. First, some typical forest plots with known features are spectrally identified. A maximum likelihood fuzzy classification with nonparametric priors is then applied to the study images, so as to derive fuzzy membership grades for all pixels with respect to the typical plots. Finally, these grades are used to compute the estimates of the forest parameters by a weighted average strategy. The method was tested on a complex, rugged area in Tuscany mainly covered by deciduous and coniferous forests. Two TM scenes and accurate ground references taken in spring and summer 1991 were utilized for the testing. The first results, statistically evaluated in comparison with those of a more usual multivariate regression procedure, are quite encouraging. The possible application of the fuzzy approach to other cases of environmental monitoring is finally discussed. >

Integration of ancillary data into a maximum-likelihood classifier with nonparametric priors

Abstract The inclusion of prior probabilities derived from the frequency histograms of the training sets has already been demonstrated to significantly improve the performance of a maximum-likelihood classifier. Based on the same principles, a method is presently proposed to integrate the information of ancillary data layers (morphology, pedology, etc.) into the classification process. The statistical basis of this probabilistic approach is described along with a procedure for the preliminary estimation of the information content expressed by the ancillary data about the cover categories. A case study is then illustrated concerning a rugged area in Tuscany (central Italy) sensed by multitemporal Landsat Thematic Mapper (TM) scenes. Ground references of nine cover categories were collected and digitized together with four ancillary data layers (elevation, slope, aspect and soils). A maximum-likelihood classification with nonparametric priors based only on the TM scenes was first tested, yielding a kappa accuracy of 0.744. The ancillary data were then analysed and integrated into the modified classifier, with notable increases in classification accuracy (up to κ = 0.910). It is concluded that the utility of such an approach must be evaluated in relation to the spectral separability of the cover categories considered and the information content of the ancillary layers.

Use of multitemporal information to improve classification performance of TM scenes in complex terrain

Abstract The discrimination of land cover types by means of satellite remotely sensed data is a very challenging task in extremely complex and heterogeneous environments where the surfaces are hardly spectrally identifiable. In these cases the use of multitemporal acquisitions could be expected to enhance substantially classification performance with respect to single scenees, when inserted in procedures which exploit all the information available. The present work discusses this hypothesis and employs three TM scenes of gently undulated terrain in Tuscany (central Italy) from different seasons of one year (February, May and August). The three phenological stages of the vegetated surfaces provided additional statistical information with respect to single scenes. Classification was tested with gaussian maximum likelihood classifiers, both separately on each of the three TM passages and, suitably adapted, on the whole multitemporal set. An iterative process using probabilities estimated from the error matrices of previous single image classifications was also tested. Results of tests show that multitemporal information greatly improves classification performance, particularly when using the statistical procedure described.

Characterization of primary productivity levels of Niger by means of NOAA NDVI variations

Abstract Usual ecoclimatic zonation schemes are based on ground collected information such as meteorological and topographic data. Such schemes are not capable of identifying all the complex environmental factors which can affect vegetation dynamics. On the other hand, the NDVI (Normalized Difference Vegetation Index) acquired by NOAA AVHRR can be a direct indicator of vegetation quantity, type and condition on a regional scale. In the present work a characterization of a Sahelian country, Niger, according to its primary productivity levels is developed based on the analysis of intra and inter‐year variations in NDVI. The mean NDVI levels of the sub‐districts from four successive rainy seasons were first compared to the relevant ground derived estimates of primary productivity. Then, by an analysis of the seasonal NDVI profiles, the identification of three main zones with different ecological characteristics (woodland, Sahel, desert) was performed. Given the enormous ecological and social importance of th...

Automatic identification of end-members for the spectral decomposition of remotely sensed scenes

Several methods have been proposed for the extraction of latent information from multispectral remotely sensed scenes based on the definition of indices and rotational transformations. A common drawback of these techniques is that they are ultimately based only on statistical relationships among pixel values rather than on physical characteristics of the scenes. Linear pixel unmixing is an alternative method which assumes that the pixel signal is the linear combination of some basic spectral components the fractions of which can be retrieved with good approximation. The method is straightforward and produces results which can be easily interpreted, but presents the problem of the identification of suitable end-members, which generally requires some external knowledge. In order to overcome this problem, in the present research a statistical method is developed for the automatic identification of end-members. This methodology is composed by several steps, that are describe and then applied to a case study with a Landsat 5 TM scene from Central Ethiopia (Africa). The results, evaluated in comparison with those of a more usual principal component transformation, indicate the good performance of the new procedure.

Multi‐scale classification of remotely sensed data by the maximization of fuzzy membership grades

Abstract Recent studies have shown that the optimal spatial resolution of remotely sensed data for the identification of scene elements is extremely variable depending on several factors. In particular, for the classification of each cover class using satellite sensor imagery there is generally an optimum aggregation level which is mainly related to the average size of the ground plots and to their internal homogeneity. In the present paper a methodology is presented for finding the relative optimum filtering size for the attribution of all pixels to some ground classes based on the maximization of fuzzy membership grades. Since these grades are indicative of the confidence of the pixel attribution to the classes, they can be used to select the best scale among differently filtered images. This hypothesis is investigated by means of a case study concerning a rugged, complex area in Central Italy sensed by bi‐temporal TM scenes. The results testify to the good performance of the approach which also shows p...