Characterisation of agricultural land using signal processing and cognitive learning techniques

This paper presents the application of neural networks and singular value decomposition to agricultural land use classification. Neural networks (NN) have been found to have good generalization properties and their use is becoming increasingly prevalent in the field of remote sensing. However, there are a number of problems where neural networks do not necessarily provide an optimum solution, these include mixed pixel analysis, sub-class characterization and parameter extraction for use in bio-physical models. Typically the application of NN techniques to remote sensing involves using one NN to classify a large number of land-cover classes. The authors have found this approach to be inefficient and inaccurate, a modular approach is therefore implemented which is more flexible. SVD has previously been used to classify agricultural species with accuracies approaching 95% and has also been used to characterise subclasses of winter wheat. SVD and key vector analysis also enable parameters such as yield to be directly correlated with the output vectors. A holistic classification process has been developed by the author's which uses the best elements from neural networks and SVD. This paper applies these techniques to optical airborne data at varying resolution from 1 m to 5 m resolution. The area used for this work is a research farm in Bavaria, Germany, which comprises of a highly dynamic terrain with small field units. High resolution land-use maps and yield data have been produced for the research farm, using GPS equipment attached to crop harvesters. These maps are used to validate the results produced by the various techniques.