Figure 8‑12 Land cover, 1986, Chiquitania

Figure 8‑13 Distance raster (metres), anthropogenic disturbance, Chiquitania

All image/raster files used in this model are co-registered 987x927 pixels or cells, where one pixel=150mx150m. Four of the seven input rasters are generated from vector data using a simple distance transform ― i.e. each pixel provides the distance to the closest object in the underlying vector set. These include urban areas, roads, streams and areas subject to anthropogenic disturbance (Figure 8‑13). The underlying dataset used in this example is from the Conservation International’s Centre for Applied Biodiversity Science, Museo Noel Kempff Mercado, Bolivia, and supplied with the Idrisi software installation. Two of the input rasters provided the elevation and slope data for the study area, and the last of the 7 inputs is another computed raster, this time providing a measure of the likelihood of individual raster cells changing landcover based on the frequency patterns found for those cells that were subject to change in the period 1986-94.

Our principal interest is in the application of the MLP modelling and classification process (Figure 8‑14) within LCM. In this instance the MLP network topology is of the form 7-7-8 (plus bias), i.e. 7 input nodes (in this instance, Bands 1, 2, 3…), all of which are real-valued raster files, 1 hidden layer with 7 nodes, and 8 output nodes.

Figure 8‑14 MLP Classifier — Idrisi

The output nodes are the 4 main land-use transitions (classes) that are taking place in the region and 4 ‘persistence classes’ ― the latter represent raster cells that might have changed between 1986 and 1994, but did not. The authors state that the inclusion of these additional output nodes assists the MLP training process. Each of the 4 main transitions observed were to the class ‘anthropogenic disturbance’ (class 8) and transferred from the identified classes: (i) woodland savannah; (ii) Amazonian rain forest; (iii) savannah; or (iv) deciduous forests. The MLP training site specification (Figure 8‑14, centre left) automatically picks a random set of cells (pixels) ― i.e. training sites ― to be used for training the model and a similar number of sites for testing the model. The number is chosen by examining the smallest transition in the group of four that are the principal subject of the modelling process. As can be seen from Figure 8‑14, key training parameters for the MLP modelling process are specified via this form. These include: using a dynamic learning rate, which commences with a value of 0.005 and reduces to 0.0001 ― this reduction takes place over the 5000 iterations or epochs of the model, but in this example most of the training was achieved within the first 1000 epochs (see the RMSE graph in Figure 8‑14); applying a momentum factor ― a value of 0.5 was specified in this example; and specifying the slope parameter ― the activation function used within Idrisi is the logistic or sigmoidal function, and here was applied with slope parameter (constant a in Figure 8‑14) of 1. Normalisation of the input raster files is not specified within the Idrisi documentation, but is assumed to be carried out by default. Three stopping criteria can be specified, although here the training stops after the maximum number of iterations has been reached. Notice that an accuracy rate of 81.5% of transitions has been obtained with this model (Figure 8‑14, running statistics panel). The authors suggest that 75% is good and 80%+ is a preferable target. This figure represents the number of correctly predicted transitions for the test pixels.

A set of four transition potential maps can then be generated using this information. One such map is shown in Figure 8‑15. The next stage in the process is to use the transition potentials map, in conjunction with Markov Chain analysis, to predict landcover in the year 2000, for which actual landcover information is known. This enables the quality of the MLP modelling and subsequent predictive process to be evaluated (see Idrisi documentation for more details of these steps in the modelling process).

Figure 8‑15 Transition potential map

Table 8‑6 W weights matrix, Chiquitania MLP model