Until recently, methodologies to extract sub-pixel information from remotely sensed data have focused on linear un-mixing models and so called fuzzy classifiers. Recent research has suggested that neural networks have the potential for providing sub- pixel information. Neural networks offer an attractive alternative as they are non- parametric, they are not restricted to any number of classes, they do not assume that the spectral signatures of pixel components mix linearly and they do not necessarily have to be trained with pure pixels. The thesis tests the validity of neural networks for extracting sub-pixel information using a combination of qualitative and quantitative analysis tools. Previously published experiments use data sets that are often limited in terms of numbers of pixels and numbers of classes. The data sets used in the thesis reflect the complexity of the landscape. Preparation for the experiments is canied out by analysing the data sets and establishing that the network is not sensitive to particular choices of parameters. Classification results using a conventional type of target with which to train the network show that the response of the network to mixed pixels is different from the response of the network to pure pixels. Different target types are then tested. Although targets which provide detailed compositional information produce higher accuracies of classification for subsidiary classes, there is a trade off between the added information and added complexity which can decrease classification accuracy. Overall, the results show that the network seems to be able to identify the classes that are present within pixels but not their proportions. Experiments with a very accurate data set show that the network behaves like a pattern matching algorithm and requires examples of mixed pixels in the training data set in order to estimate pixel compositions for unseen pixels. The network does not function like an unmixing model and cannot interpolate between pure classes.

Thesis (MSc. (Geography)) --University of Limpopo, 2018 / Water hyacinth (Eichhornia crassipes) is recognised as the most notorious invasive species the world-over. Although its threats and effects are fully documented, its distribution is not yet understood, especially in complex environments, such as river systems. This has been associated with the lack of accurate (high spatial resolution) and robust techniques, together with the reliable data sources necessary for its quantification and monitoring. The advent of new generation sensors i.e. Landsat 8 Operational Land Imager (OLI) and Sentinel-2 MultiSpectral Instrument (MSI) data, with unique sensor design and improved sensing characteristics is therefore perceived to provide new opportunities for mapping the distribution of invasive water hyacinth in small waterbodies. This study aimed at mapping and understanding the spatio-temporal distribution of invasive water hyacinth in the Greater Letaba river system in Tzaneen, Limpopo Province of South Africa using Landsat 8 OLI and Sentinel-2 MSI data. Specifically, the study sought to identify multispectral remote sensing variables that can optimally detect and map invasive water hyacinth. Landsat 8 OLI and Sentinel-2 MSI were tested based on the spectral bands, vegetation indices, as well as the combined spectral bands plus vegetation indices, using discriminant analysis algorithm. From the findings, Sentinel-2 MSI outperformed Landsat 8 OLI in mapping water hyacinth, with an overall classification (OA) accuracy of 77.56% and 68.44%, respectively. This observation was further confirmed by a t-test statistical analysis which showed that there were significant differences (t=6.313, p<0.04) between the performance of the two sensors. Secondly, the study sought to map the spatial distribution of invasive water hyacinth in the river system over time (Seasonal). Multi-date 10 m Sentinel-2 MSI images were used to detect and monitor the seasonal distribution and variations of water hyacinth in the Greater Letaba River system. The study demonstrated that, about 63.82% of the river system was infested with water hyacinth during the wet season and 28.34% during the dry season. Sentinel-2 MSI managed to depict species spatio-temporal distribution with an OA of 80.79% during wet season and 79.04% in dry season, using integrated spectral bands and vegetation indices. New generation sensors provide new opportunities and potential for seasonal or long-term monitoring of aquatic invasive species like water hyacinth- a previously challenging task with broadband multispectral sensors. / Risk and Vulnerability Science Centre (RSVC)

La thèse propose des nouvelles techniques pour la classification et le démelange spectraldes images obtenus par télédétection iperspectrale. Les problèmes liées au données (notammenttrès grande dimensionalité, présence de mélanges des pixels) ont été considerés et destechniques innovantes pour résoudre ces problèmes. Nouvelles méthodes de classi_cationavancées basées sur l'utilisation des méthodes traditionnel de réduction des dimension etl'integration de l'information spatiale ont été développés. De plus, les méthodes de démelangespectral ont été utilisés conjointement pour ameliorer la classification obtenu avec lesméthodes traditionnel, donnant la possibilité d'obtenir aussi une amélioration de la résolutionspatial des maps de classification grace à l'utilisation de l'information à niveau sous-pixel.Les travaux ont suivi une progression logique, avec les étapes suivantes:1. Constat de base: pour améliorer la classification d'imagerie hyperspectrale, il fautconsidérer les problèmes liées au données : très grande dimensionalité, presence demélanges des pixels.2. Peut-on développer méthodes de classi_cation avancées basées sur l'utilisation des méthodestraditionnel de réduction des dimension (ICA ou autre)?3. Comment utiliser les differents types d'information contextuel typique des imagés satellitaires?4. Peut-on utiliser l'information données par les méthodes de démelange spectral pourproposer nouvelles chaines de réduction des dimension?5. Est-ce qu'on peut utiliser conjointement les méthodes de démelange spectral pour ameliorerla classification obtenu avec les méthodes traditionnel?6. Peut-on obtenir une amélioration de la résolution spatial des maps de classi_cationgrace à l'utilisation de l'information à niveau sous-pixel?Les différents méthodes proposées ont été testées sur plusieurs jeux de données réelles, montrantresultats comparable ou meilleurs de la plus part des methodes presentés dans la litterature. / The thesis presents new techniques for classification and unmixing of hyperspectral remote sensing data. The main issues connected to this kind of data (in particular the huge dimension and the possibility to find mixed pixels) have been considered. New advanced techniques have been proposed in order to solve these problems. In a first part, new classification methods based on the use of traditional dimensionality reduction methods (such as Independent Component Analysis - ICA) and on the integration of spatial and spectral information have been proposed. In a second part, methods based on spectral unmixing have been considered to improve the results obtained with classical methods. These methods gave the possibility to improve the spatial resolution of the classification maps thanks to the sub-pixel information which they consider.The main steps of the work are the following:- Introduction and survey of the data. Base assessment: in order to improve the classification of hyperspectral images, data related problems must be considered (very high dimension, presence of mixed pixels)- Development of advanced classification methods making use of classic dimensionality reduction techniques (Independent Component Discriminant Analysis)- Proposition of classification methods exploiting different kinds of contextual information, typical of hyperspectral imagery - Study of spectral unmixing techniques, in order to propose new feature extraction methods exploiting sub-pixel information - Joint use of traditional classification methods and unmixing techniques in order to obtain land cover classification maps at a finer resolutionThe different methods proposed have been tested on several real hyperspectral data, showing results which are comparable or better than methods recently proposed in the literature.