Skip to Main Content
Mixed pixels are one of the largest sources of error and uncertainty in mapping from remotely sensed data. A Hopfield neural network based approach to super-resolution mapping has become popular for mapping at a sub-pixel scale, partly because it seeks to maintain the class proportional information indicated by a soft classification analysis. The use of the approach is, however, handicapped by a lack of guidance on the parameter setting values and of the impacts of different landscape patterns on the analysis. Here, the sensitivity of the Hopfield neural network for super-resolution mapping is investigated with a focus on the effect of different landscape types and parameter settings using simulated and real data sets. It is shown that the method's suitability varies between landscapes, being most suited to situations in which landscape patches are large (>; 1 pixel) . Additionally, for such landscapes the widely used scenario in which the weighting parameters are set at equal values is successful but the approach is less effective for the mapping of small isolated land cover patches. With the latter, it is shown to be important to weight the area constraint highly and undertake a large number of iterations. Critically, it is shown that equal weighted parameter settings and imbalanced settings to emphasize the area constraint are most suitable for landscapes comprising large and small patches respectively. Moreover, the positive attributes of these two sets of parameter settings may be combined to yield an enhanced mapping method for landscapes that comprise a mixture of patch sizes.