A Neural Network Approach For Interpolating Species Density Patterns From Remotely Sensed & GIS data: An Example Using The Desert Tortoise

About The Authors

Abstract

A backpropagational neural network simulator provides field ecologists with a statistically-based approach for correlating field measurements with remotely sensed digital imagery, digital elevation models, landscape thematic maps and other raster Geographic Information System (GIS) map coverages. Once trained, the neural network is applied to each raster grid cell of the digital landscape to yield a two-dimensional map showing the interpolation of the field data to every grid cell. The result is a digital image which is shown to yield a rich spatial data set based on ecological field measurements. The example demonstrated in this document correlates field transect data of desert tortoise surveys with 21 digital GIS map layers of environmental and remote sensed spectral attributes.

Study Problem

Landscape ecologists are utilizing spatially explicit models to design, develop, and test concepts of ecosystem and landscape dynamics. One approach is to devide a heterogeneous landscape into a systematic pattern of grid cells. Each cell is treated as a homogeneous system and is simulated with a single ecosystem model. The model is applied simultaneously to each cell. Cells are characterized by their individual internal states and the states of their immediated neighbours. Such models must be initialized by providing the status of every one of the system's state variables on a landscape scale. The problem is that spatially explicit ecological parameters (e.g. distribution of populations and communities) are typically unavailable for full landscape coverages. The challenge that this demonstration addresses concerns how to efficiently convert transect data sampled from habitat patches to full-coverage landscape data.

This study was conducted to support the initialization of a dynamic spatial ecological model of the desert tortoise. This model requires initialization with tortoise densities for each grid cell in the model. A back-propagation neural network was used to associate tortoise survey field data with available digital raster GIS map layers representing environmental and habitat attributes. This demonstration presents the basics of a neural network technique only. A forthcoming academic publication will more fully evaluate this work in the context of the two conventional interpolation approaches, such as linear regression and thin-plate splines.

Approach

According to Rumelhart et al. (1986) a neural network generally consists of the following components: The network topology and the form of the rules and functions are all learning variables in a neural network learning system, leading to a wide variety of network types. Some of the well known types of neural networks are: Competitive Learning (Grossberg 1976; Zipser 1985), the Boltzmann Machine (Hinton 1984), the Hopfield Network (Hopfield, 1982), the Kohonen network (Kohonen 1988), the Adaptive Resonance Theory (ART) (Grossberg 1987), and back propagation neural networks (Rumelhart 1986). Although there are many other variations of neural networks, the back propagation network and its variants, as a subset of multilayer feed forward networks, are currently the most widely used networks in applications.

Back propagation neural networks are loosely based on the neuronal structure of the brain and provide a powerful statistical approach for exploring solutions of non-linear systems (Rumelhart 1986). This study employs a back propagational neural network which was used to correlate input information with matched output values. The solution space is the set of all combinations of coefficients for the equations that are being used for modeling a system of interest.

A neural network starts with a set equation associated with a randomly chosen group of coefficients and rules for adjusting those coefficients during "training". The solution space is explored using a gradient descent method which converges on the global minimum; the set of equation coefficients which provides the absolute best solution (mapping of inputs to outputs) using the pre-determined equations.

Neural Networks are analogous to Monte Carlo simulations and genetic algorithms, and are used to explore a system or set of data for the purpose of finding a solution that has predictive capabilities. Back propagation neural networks are utilized when a set of inputs are known to match corresponding outputs, but the nature of the relationship is unknown. A network attempts to find a mathematical relationship between the defined inputs and outputs as it is repetitively presented with the data through a supervised "training" process.

Supervised training is analogous to the learning behavior of a child's mind as it is presented with samples of items from different categories along with the correct interpretation of each sample; for example, learning to recognize dogs and cats from examples of each. And, like the learning child, the trained network does not yield logical "reasons" for differentiating. They are simply 'known' to be different. A neural network can yield a set of coefficients (weights), but provides no logical descriptions, cause-effect relationships, or expert system rules.

The neural network software package, "BP" (for backpropagation, written by James Westervelt, USACERL) was utilized by this project. BP was originally intended for use in association with the GIS system GRASS, which is a full-feature public domain GIS system written by the Army Corps of Engineers. GRASS was used in this research to manipulate and display all spatial data. The two-way link between the neural network software and the GIS is a cr itical component to this work, and the authors expect 'bp' and programs like it to become included as analytical tools built into future GIS and image analysis systems.

Network Design Parameters

Employing a backpropagational neural network requires an understanding of a number of network design options. A general background of neural networks is given elsewhere in this document, however a brief discussion of some key network parameters is given below. Be advised that there are no definate rules for choosing the settings of these parameters a priori. Since the solution space associated with each problem is not known, an number of different network runs must be undertaken before the user can determine with relative confidence a suitable combination.