Read Papers

GEOBIA, 2010.

Doxani, G., Karantzalos, K., Tsakiri-Strati, M.

The work presents and object-based classification framework to detect building constructions using two different times of remote sensing imagery. The method applies the Multivariate Alteration Detection (MAD), which is a correlation analysis between two groups of variables (images). The obtained MAD components are the difference of the corresponding input images, which depict the same area and were acquired at different dates.

MAD components with values higher than a certain threshold |two times the standard deviation| correspond to changing pixels. Such pixels are then analysed further to infer changes regarding building constructions. They tested using single pixels, and also the spectral information obtained by segmentation in different scales.

MultiTemp, 2003.

Heas, P., Datcu, M., Giros, A.

This work performs two types of classification using image time-series. The first one is called time-localized clustering, where each time-window in the series possesses a corresponding classification. Such classification neglects the information about the causalities between different times. The second is called multi-temporal classification, where information from all images are used to perform the classification.

The authors argue that it is possible to create a model which is capable of measuring the distance between the multi-temporal clusters and the time-localized clusters. With this model it is possible to trace the cluster evolutions. They calculate the cross entropy between each multi-temporal cluster and all the time-localized ones. The maximum entropy is used to match both types of clusters.

Advances in Conceptual Modeling, 2010.

Hornsby, K., Egenhofer, M.J., Hayes, P.

This paper describes a formal way to represent cyclic events. Certain changes in the environment occur cyclically, and current information systems do not deal with it. They proposed a representation of 16 different relations between two events which occur in cycles, representing they as arcs in a circle.

According to the article, the most general model of time in a temporal logic represents time as an arbitrary, partially-ordered set. In the linear model, there is total order on time, resulting in the linear advancement of time from the past, through the present, and to the future. The branching model describes time as being linear from the past to the present, and then divides into several time-lines, each representing a potential sequence of events. However, both models do not treat the fact that certain events or phenomena may be recurring.

Article definitions on temporal data models:

• time points: typically describe a precise time when an event occurred;
• time intervals: used when precise information on time is unavailable. Much of our temporal knowledge is relative and methods are needed that allow for significant imprecision in reasoning

ACM SIGMOD, 2004

Faloutsos, C., Ranganathan, M., Manolopoulos, Y.

This work investigates a technique to match subsequences of values in time-series of values. A time-series is described as an 1-dimensional sequence of values. To avoid brute force search on every sequence for the subsequences, they propose to apply Discrete Fourier Transform in the data, and keep only the first few coefficients, called features. With these features, rectangles in the feature space are created to join pieces of the values in a smaller representation.

These rectangles are created using subsequences of the entire databases, with the same length of the query sequence. Then the search for matchings is performed in these new space, reducing processing time.

IGARSS, 2004.

Heas, P., Datcu, M., Giros, A.

This paper introduces the concept of two different representations for image time series:

1. “The spatio-temporal representation is simply the time-series of images.”
2. “The multitemporal feature space representation is a multidimensional space composed by the union of all the time localized feature components and for which the spatial index is hidden.”

The proposal is to search for similar trajectories in time series. The proposed method is based on Mutual Information, which quantifies the amount of information that is contributed by one random variable into another. According to the paper, “mutual information evolutions are significant of the degree of change of the multitemporal cluster during time. Set of consecutive nodes can be grouped by similarity according to this measurement.” The detection of similarities in the oriented graphs is done by discovering the probability density function (PDF) of the nodes, and measuring the entropy of these functions, also considering the irregular time sampling rate.

ACM SIGKDD international conference on knowledge discovery and data mining, 2008.

Boriah, S., Kumar, V., Steinbach, M., Potter, C., Klooster, S.

This paper proposes a new technique for change detection in satellite image time series. Change detection has been already studied in many research areas, such as statistics, signal processing, and control theory. However, the employed techniques in these areas are not well suited to huge amount of data present in Earth Science. In this article they analyzed a vegetation-related variable, “the enhanced vegetation index (EVI) product, measured by MODIS. EVI is a vegetation index which essentially serves as a measure of the amount and 'greenness' of vegetation at a particular location.”

The proposed technique for change detection is called Recursive Merging, and is based on some premises:

• Given the large coverage of land cover data sets, it is fairly obvious that only a small fraction of points will actually exhibit a change.
• Natural seasonal growing cycle is a dominant characteristic of a time series and this intrinsic seasonality should not itself be called a change.
• The main idea is to exploit seasonality in order to distinguish between points that have had a land cover change and those that have not.

The algorithm runs as follows:

1. The two most similar consecutive annual cycles are merged, and the distance is stored.
2. Step 1 is applied recursively until one annual cycle is left remaining.
3. The change score for this location is based on whether any of the observed distances are extreme, by calculating the ratio between the minimum and the maximum distances.

If the change score is greater than a certain threshold, the point is then classified as a change.

Definition of land cover change, by the authors: “The land cover change detection problem is essentially one of detecting when the land cover at a given location has been converted from one type to another”.

IGARSS, 2008.

Andreea, J., Meger, N., Trouve, E., Bolon, P.

The authors use Satellite Image Time Series (SITS) to extract pixel-based evolutions and to detect spatio-temporal patterns in the images. The first technique is based on pixel quantization. Pixel values are quantized in a set of non-overlapping intervals equally populated. This reduces the amount of values, and therefore the evolution is described as a sequence of quantized values along the time. If using more than one image channel, for each time is created a sequence of values. To this sequence of values is then aggregated the subsequent snapshots, and the final sequence defines the evolution. Then, “pixels having the same evolution at the same dates are then set to the same color”.

The quantization reduces the amount of data, and therefore reduces the computational time to reach results, however, it may produce only rough results. Another drawback of this technique is that evolutions are only classified as similar if they occurred in the same interval, but sometimes the same event can occur in different intervals, and still should be classified as the same evolution.

ACM symposium on Applied computing, 2008.

Mota, J., Câmara, G., Fonseca, L., Escada, M., Bittencourt, O.

Starting from one initial pattern of deforestation, they aim to recover the object's history, based on a set of similar cases of deforestation progress. For this purpose they use a Case-Based system, which stores previous cases, and “continually increases allowing adapting cases with similar characteristics that can be useful to solve a new problem”.

The database of cases were generated in the Agrarian Settlements Project called Vale do Anari, in Rondônia. The main spatial patterns are called Linear, Irregular, and Geometric. The stages of deforestation were defined as Road, Small Lots, and Large Concentration. Such typology were applied to describe the deforestation evolution in another area.

Remote Sensing of Environment, 2005.

Xu, M., Watanachaturaporn, P., Varshney, P. Arora, M.

This work proposes a decision tree with soft thresholds, which decompose the pixel into its class constituents in the form of class proportions. The outputs from soft classification provide a set of fraction images that represent proportion of classes for each pixel. The use of decision trees is justified because remote sensing data often not follow Gaussian distribution, an assumption made by several classifiers.

To build the class proportions, the authors suggest to create several decision trees, called forest. Every class proportion will derive one soft decision tree in the forest. Concluding, the authors argue that the tree construction is very fast and is, therefore, suitable for classifying large remote sensing images.

IEEE TGRS, VOL. 37, NO. 2, MARCH 1999

Friedl, M., Brodley, C., Strahler, A.

The authors evaluated the boosting technique using decision trees algorithm to land cover classification. The boosting technique estimate multiple classifications iteratively. At each iteration, a weight is assigned to each training observation. Wrong classifications in the previous iteration get a bigger weight in the next iteration, “forcing” the classification algorithm to concentrate on more difficult observations to classify.

They concluded that boosting is a useful technique and should be used for land cover classification problems using remotely sensed data at continental to global scales. Besides, the use of geographic position provides substantial predictive power to the decision tree classification algorithms. The proved this assumption by classifying fairly coarse classes of vegetation at continental and global scales. However, geographic position should only be used as a secondary input feature used to discriminate between land cover classes that are spectrally similar, but geographically distinct.

Remote Sensing of Environment, 2010.

Powell, S., Cohen, W., Healey, S., Kennedy, R., Moisen, G., Pierce, K., Ohmann, J.

The authors modeled tree biomass in a 20 year time series of Landsat images (northern Arizona and northern Minnesota) and tested three statistical techniques to derive trajectories of evolution.

1. Reduced Major Axis regression: an orthogonal regression technique that aims to minimize error in both X and Y directions.
2. Gradient Nearest Neighbor imputation …
3. Random Forests regression trees: RF is a non- parametric ensemble modeling approach that constructs numerous small regression trees that vote on predictions, and is considered to be robust to over-fitting.

The following variables were employed to model live aboveground tree biomass:

• Raw Landsat bands (B1–B5, B7, as surface reflectance)
• Tasseled Cap indices (brightness (TCB), greenness (TCG), and wetness (TCW))
• NDVI, ( ( B4 − B3 ) / ( B4 + B3 ) )

The results obtained from the three techniques were compared to a validation data set, in terms of root mean square error (RMSE), variance ratio (VR) and bias (difference between observed and predicted values). According to the authors, in terms of RMSE, the technique of RF obtained the best results. Therefore, if minimizing the prediction error is the main objective, RF is suggested. Another conclusion is that “although there is significant modeling error with biomass prediction, the temporal analysis ensures that at least the models are consistent across the time-series, and, therefore, the relative changes are potentially accurate”.

International Journal of Remote Sensing, 2008.

Wang, Y., Li, J.

Decision Tree (DT) was tested as a feature selection algorithm, using hyperspectral data. Feature selection is defined by the authors as imperative for massive amounts of data. According to the authors, “feature selection results of DT are those features that are used to form splitting rules at internal tree nodes”, i.e. the attributes of the splitting nodes are the most relevant ones. C4.5 algorithm were employed in this article.

DT as a feature selector is proportional to the amount of samples (“tends to select too many features when the sample size is large”), and they suggested to use another algorithm to select features (mainly with small sample sizes), and to allow repeated use of features to form split nodes.

Another suggestion is to use “feature extraction”, which means creating new features by applying combinations on already existent features. However, “although extracted features have higher discrimination power, their physical meanings were hard to explain, which lowered the interpretability of classification trees”.

Remote Sensing of Environment, 2003.

Small, C.

The author explores the use of IKONOS imagery for intraurban classification. Medium spatial resolution sensors, such as TM and SPOT present abundance of mixed pixels. However, mixed pixels are problematic for statistical classification methods because most algorithms are based on the assumption of spectral homogeneity at pixel scale within a particular class of land cover, but urban areas provide examples of spectrally diverse, scale-dependent thematic classes containing large numbers of pixels that are spectrally indistinguishable from other land cover classes. The work proposes the use of spatial autocorrelation to quantify the characteristic scale lengths of urban reflectance within and among different cities. The question of the paper is if the consistency of the spectral mixing space for a urban areas can be provided by a simple three-component linear mixture model, to characterize the urban reflectance. Concluding, it is argued that spectral mixture analysis is preferable to “hard classification” for many applications because it accommodates the preponderance of mixed pixels observed in almost all multispectral imagery.