C.A.J.M. de Bie

LaHMa: a landscape heterogeneity mapping method using hyper-temporal datasets

A new quantitative method extracts a landscape heterogeneity map (LaHMa) from hyper-temporal remote-sensing data. The feature extraction method is data-driven, unbiased, and builds on the commonly used data reduction technique of Iterative Self-Organizing Data Analysis (ISODATA) clustering with the support of divergence separability indices. First, the relevant spatial–temporal variation in normalized difference vegetation index (NDVI) is classified through ISODATA clustering. Second, a series of prepared cluster maps are overlaid to examine and detect the frequency with which boundaries between clusters occur at the same location. This step identifies the boundary strength between clusters and detects spatial heterogeneity within them. Results of the method are explored for the typical agriculture-defined landscape of the Mekong delta, Vietnam, using NDVI-imagery time-series from SPOT-Vegetation and MODIS-Terra. The method extracts useful landscape heterogeneity features and can support land-cover mapping requiring information on fragmentation and land-cover gradients.

Hyper-temporal SPOT-NDVI dataset parameterization captures species distributions

Hyper-temporal SPOT NDVI images contain useful information about the environment in which a species occurs, including information such as the beginning, end, peak, and curvature of photosynthetically active vegetation (PAV) greenness signatures. This raises the question: can parameterization of hyper-temporal SPOT NDVI images be useful to predict species distribution? A set of SPOT-NDVI images for the whole of Ethiopia covering nine years was classified using the unsupervised ISODATA clustering algorithm to group similar NDVI pixel values. The HANTS (Harmonic ANalysis of Time Series) algorithm, that fits series of smoothing cosine waves, was then applied to the time series for each of the NDVI classes to generate seven output Fourier components. These components, together with the topographic parameters slope and elevation, were used as predictors in a species distribution model using MAXENT. Presence-only data of one test species, Boswellia papyrifera, were modelled. This species is diminishing at an alarming rate and requires conservation. The performance of the model was evaluated by the area under curve (AUC) of the receiver-operating characteristics value. The output distribution map was tested for its agreement with the NDVI-clustering approach and conventional B. papyrifera distribution map using Kappa. The relative contributions of the first four predictors to the MAXENT in sequence were: 2nd harmonic phase, elevation, amplitude of the 1st harmonics, and amplitude of the 2nd harmonics. The average AUC test result for the 100 runs was 0.98 with a standard deviation of 0.002. The probability distribution map clearly shows high correlation with the B. papyrifera occurrence data. In addition, the distribution map was found to be in agreement with the NDVI-clustered and conventional map with improved details. Classifying hyper-temporal NDVI images and extracting their parameters through the use of the HANTS algorithm captures the PAV greenness behaviour (parameters) of the environment of the species studied. These parameters have proved successful in predicting the distribution of B. papyrifera.

A common dominant scale emerges from images of diverse satellite platforms using the wavelet transform

In this article we investigate the scale dependence of spatial heterogeneity in multiresolution and multisensor data using the wavelet transform. The landscape analysed with the wavelets retains the same dominant pattern irrespective of the original pixel size of the image. In agricultural areas, typically characterized by a mosaic of cultivated fields, the wavelet transform quantified consistently a median dominant scale of 512 m in the Orthophoto, Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and Landsat Enhanced Thematic Mapper Plus (ETM + ). The dominant scale represented the dominant field size of cultivated areas. The shape of the average wavelet energy curves was also similar among the images. In semi-natural areas the wavelet transform quantified consistently a median dominant scale of 128 m in the Orthophoto and ASTER. The median dominant scale of ETM +  was slightly smaller and located at 64 m. We characterized the spatial heterogeneity of agricultural and semi-natural areas in Andalucía (Spain) using multisensor data not time coincident ranging from 1 m (Orthophoto), to 15 m (ASTER) to 28.5 m (ETM + ). The contrast in vegetation cover was measured using Normalized Difference Vegetation Index (NDVI) in ASTER and ETM +  and red band in Orthophoto. We performed a multiresolution analysis using a Haar two-dimension discrete wavelet transform to quantify and compare the intensity (maximum degree of contrast in vegetation cover), the dominant scale (the scale at which the maximum intensity occurs) and the wavelet energy curve (intensity plotted as a function of the scale) of different images at intervals of power of 2 within the scale range from 2 to 4096 m.