R. De Wulf

Using genetic algorithms in sub-pixel mapping

In remotely sensed images, mixed pixels will always be present. Soft classification defines the membership degree of these pixels for the different land cover classes. Sub-pixel mapping is a technique designed to use the information contained in these mixed pixels to obtain a sharpened image. Pixels are divided into sub-pixels, representing the land cover class fractions. Genetic algorithms combined with the assumption of spatial dependence assign a location to every sub-pixel. The algorithm was tested on synthetic and degraded real imagery. Obtained accuracy measures were higher compared with conventional hard classifications.

Monitoring Sahelian floodplains using Fourier analysis of MODIS time-series data and artificial neural networks

Fourier analysis of Moderate Resolution Image Spectrometer (MODIS) time‐series data was applied to monitor the flooding extent of the Waza‐Logone floodplain, located in the north of Cameroon. Fourier transform (FT) enabled quantification of the temporal distribution of the MIR band and three different indices: the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Water Index (NDWI), and the Enhanced Vegetation Index (EVI). The resulting amplitude, phase, and amplitude variance images for harmonics 0 to 3 were used as inputs for an artificial neural network (ANN) to differentiate between the different land cover/land use classes: flooded land, dry land, and irrigated rice cultivation. Different combinations of input variables were evaluated by calculating the Kappa Index of Agreement (KIA) of the resulting classification maps. The combinations MIR/NDVI and MIR/EVI resulted in the highest KIA values. When the ANN was trained on pixels from different years, a more robust classifier was obtained, which could consistently separate flooded land from dry land for each year.

A methodology for the description of border hedges and the analysis of variables influencing their distribution: a case study in western Kenya

This paper presents results from a survey of border hedges on farmland in western Kenya. The survey covered 160000 ha of high potential land in eastern Siaya District and Vihiga District of western Kenya. The survey attempted to widen the knowledge of the typology, the biomass and the parameters influencing the spatial distribution of hedge types. Spatial analysis was used to delimit hedge type sub-regions (using cluster analysis) and to identify the variables influencing the spatial distribution of hedge types (using discriminant analysis). It is demonstrated that a complex association of variables is influencing the subdivision of the two districts in hedge type sub-regions in which ethnicity, population density, area in woodlots and ecological variables like elevation, rainfall and soil fertility are important variables. These variables are influencing each other and are responsible for the contrasting situation in Vihiga and Siaya District. Border hedges have similar functions in both districts (demarcation of land, to prevent cattle from entering), nevertheless species composition and dimensions differ remarkably in both districts. Border hedges in Siaya District are poorly managed or not managed at all. In Vihiga District people are used to manage their hedges. Agroforestry techniques, for example techniques based on frequent pruning of border hedges, have a high chance in being successful in this district because no additional investment in labour or time is required. The spatial distribution in the amount of biomass is strongly correlated with the distribution in the per cent area ground cover of border hedges. This means that secondary data on the area in hedges derived from aerial photographs can serve as a useful indicator of the biomass present. As a result, the most difficult part of the field survey, the destructive sampling for the determination of the biomass, can be eliminated, making general surveys considerably easier.

Modelling yearly flooding extent of the Waza-Logone floodplain in northern Cameroon based on MODIS and rainfall data

The Sahelian floodplains are of high ecological and economical importance, providing water and fresh pasture in the dry season. A spatial model is presented to predict the yearly flooding extent of the Waza-Logone floodplain based on cumulative runoff in the catchment area and estimations of the soil moisture prior to the flooding. Observations of flooding extent were based on thresholding 16-day composite Moderate Resolution Imaging Spectroradiometer (MODIS) shortwave infrared (SWIR) images. The Soil Conservation Service Curve Number (SCS-CN) method was used to calculate cumulative runoff within the Logone catchment area based on rainfall estimates (RFEs) for Africa. MODIS SWIR images acquired prior to the flooding were used as indicators for soil moisture. The mean observed flooding extent of the Waza-Logone floodplain during the period 2000–2005 was 6747 km2 with a standard deviation of 1838 km2. Multiple regression analysis was performed to create a predictive model forecasting flooding extent 1.5 months in advance with a coefficient of determination (R 2) equal to 0.957. Multiple regression modelling was also performed for three subregions separately. The 1.5-month forecast model for the Waza subregion resulted in the highest accuracy (R 2 = 0.950). A floodwater distribution map was created for this subregion model, allowing determination where the flooding occurs for an estimated flood size. The average additional error caused by the mapping procedure was 138 km2, which is relatively small compared to an average flooded area of 3211 km2 for the Waza subregion. As the flooding extent in the Waza-Logone floodplain is highly correlated to the amount of natural resources available in the dry season, the model may be a valuable tool for sustainable management of these resources.

Mapping dominant vegetation communities at Meili Snow Mountain, Yunnan Province, China using satellite imagery and plant community data

Mapping dominant vegetation communities is important work for vegetation scientists. It is very difficult to map dominant vegetation communities using multispectral remote sensing data only, especially in mountain areas. However plant community data contain useful information about the relationships between plant communities and their environment. In this paper, plant community data are linked with remote sensing to map vegetation communities. The Bayesian soft classifier was used to produce posterior probability images for each class. These images were used to calculate the prior probabilities. One hundred and eighty plant plots at Meili Snow Mountain, Yunnan Province, China were used to characterize the vegetation distribution for each class along altitude gradients. Then, the frequencies were used to modify the prior probabilities of each class. After stratification in a vegetation part and a non-vegetation part, a maximum-likelihood classification with equal prior probabilities was conducted, yielding an overall accuracy of 82.1% and a kappa accuracy of 0.797. Maximum-likelihood classification with modified prior probabilities in the vegetation part, conducted with a conventional maximum-likelihood classification for the non-vegetation part, yielded an overall accuracy of 87.7%, and a kappa accuracy of 0.861.

Semi-automated forest stand delineation using wavelet based segmentation of very high resolution optical imagery

Stand delineation is one of the cornerstones of forest inventory mapping and a key element to spatial aspects in forest management decision making. Stands are forest management units with similarity in attributes such as species composition, density, closure, height and age. Stand boundaries are traditionally estimated through subjective visual air photo interpretation. In this paper, an automatic stand delineation method is presented integrating wavelet analysis into the image segmentation process. The new method was developed using simulated forest stands and was subsequently applied to real imagery: scanned aerial photographs of a forest site in Belgium and ADS40 aerial digital data of an olive grove site in Les Beaux de Provence, France. The presented method was qualitatively and quantitatively compared with traditional spectral based segmentation, by assessing its ability to support the creation of pure forest stands and to improve classification performance. A parcel/stand purity index was developed to evaluate stand purity and the expected mapping accuracy was estimated by defining a potential mapping accuracy measure. Results showed that wavelet based image segmentation outperformed traditional segmentation. Multi-level wavelet analysis proved to be a valuable tool for characterizing local variability in image texture and therefore allowed for the discrimination between stands. In addition, the proposed evaluation measures were found appropriate as segmentation evaluation criteria.

Previously trained neural networks as ensemble members: knowledge extraction and transfer

The use of Artificial Neural Networks (ANNs) for the classification of remotely sensed imagery offers several advantages over more conventional methods. Yet their training still requires a set of pixels with known land cover. To increase ANN classification accuracy when few training data are available, an algorithm was applied that allows experience gained in previous classifications to be reused. The proposed method was evaluated by classifying a tropical savannah region in northern Togo using Landsat Thematic Mapper (TM) imagery. The presented approach reached a mean kappa coefficient that was significantly larger (at the 95% level) than that obtained after training networks with randomly initialized weights. Also, the observed variances on the obtained accuracies were significantly lower when compared to networks that were randomly initialized. Finally, Bhattacharyya (BH) distances were used to explain why some land cover classes benefit more from knowledge transfer than others.

Probabilistic Models for Linkage Learning in Forest Management

Today, forest management has become an arduous task. Forests are managed for efficient timber production, which demands large uniform stands, as well as for conservation and recreation, which require a pattern of smaller, more diverse stands. The forest management problem can therefore be regarded as an optimal patch design problem. In this chapter, the potential use of probabilistic models for linkage learning is investigated in the field of optimal patch design. The following hypothesis is investigated: linkage learning helps to solve the forest management problem and results in significantly better solutions. Two linkage learning algorithms and a simple genetic algorithm are compared and possible differences are explained in the context of this optimization problem.

Training neural networks on artificially generated data: a novel approach to SAR speckle removal

A neural network-based method for speckle removal in synthetic aperture radar (SAR) images is introduced. The method rests on the idea that a neural network learning machine, trained on artificially generated input–target couples, can be used to efficiently process real SAR data. The explicit plus-point of the method is that it is trained with artificially generated data, reducing the demands put on real input data such as data quality, availability and cost price. The artificial data can be generated in such a way that they fit the particular characteristics of the images to be denoised, yielding case-specific, high-performing despeckling filters. A comparative study with three classical denoising techniques (Enhanced Frost (EF), Enhanced Lee (EL) and Gamma MAP (GM)) and a wavelet filter demonstrated a superior speckle removal performance of the proposed method in terms of quantitative performance measures. Moreover, qualitative evaluation of the despeckled results was in favour of the proposed method, confirming its speckle removal efficiency.

Estimating Fuzzy Membership Function Based on RMSE for the Positional Accuracy of Historical Maps

When combining data from a time period spanning 300 years, these data will originate from various sources, and will be created by a diversity of techniques. Different spatial data types have different sources of error and provide information at varying accuracy levels. Some changes, detected by comparison of two misregistered maps, are caused by a difference in positional accuracy and have little to do with actual changes. This paper concentrates on this positional error, expressed as a root-mean-square error (RMSE). Due to this positional error, the actual position of the boundaries of for instance forest polygons, is uncertain. The possibility of actually finding forest increases from zero to one within a band around the forest boundaries. The aim of this paper was to analyze the relation between this possibility and the distance to the forest boundary. Using a simulation approach, distorted maps with different RMSE values were created. The superposition of these distorted maps was used to fit several functions. This was done in order to determine an appropriate form of the forest membership function. A sigmoid or a polynomial function yielded the best results, especially if the distortion was low. followed by an inverse difference weighted function. The linear function resulted in the poorest fit. The parameters of a given function type varied when the forest pattern or the RMSE value varied.