Hypertemporal image analysis for crop mapping and change detection

Many authors explored the use of multi-temporal images, recorded within a season or across years, for (i) ecosystem monitoring, (ii) land cover (crop) identification, and (iii) change detection (Copin et.al, 2004). Temporary trajectory analysis, drawing on time-profile-based data originating from a large number of observation dates, has mainly been done through threshold-based methods, compositing-algorithms, or Fourier series approximation. This paper presents findings of a multivariate change detection method that processes the full dimensionality (spectral and temporal) of 10-day composite (1998-onwards) 1-km resolution SPOT-Vegetation NDVI images. Using the ISODATA clustering algorithm of Erdas-Imagine software and all available NDVI image data layers, unsupervised classification runs were carried out. These produced minimum- and average-divergence statistical indicators that in turn were used to identify the optimum number of classes that best suited the data put to the unsupervised classification algorithm. The selected classified map is linked to a set of time-profile-based signatures (profiles) that form the map legend. Studies were carried out for (i) Portugal to identify the extend and nature of land cover units, (ii) the Limpopo valley, Mozambique to map gradients, (iii) the Limpopo valley, Mozambique, to monitor flooded areas, (iv) Garmsar, Iran, to detect spatial differences in water availability, (v) Nizamabad, India, to link NDVI profiles to land use classes and (vi) Andalucía, Spain to disaggregate reported agricultural crop statistics to 1x1km pixel crop maps. Results compose of statistical findings underpinning the method, maps showing the spatial-temporal characteristics of the findings, and the applicability of the method for the studied topics.