L'analyse de la Polarimétrie SAR pour la détection des structures archéologiques de surface et de subsurface du site de Gebel Barkal (Sudan), inscrit dans la Liste du Patrimoine Mondial depuis 2003, est l'objectif des travaux de recherche effectués dans le cadre de cette thèse de doctorat. En particulier, les capacités de pénétration dans le sol des bandes C et L ont été analysées grâce à l'utilisation des images des capteurs ALOS PALSAR (archivées) et RADARSAT-2 (spécifiquement acquises). En outre, l'activité de recherche illustre les potentialités de l'intégration des données satellitaires polSAR et optiques dans un projet SIG dédié, réalisé grâce à une collaboration avec les Universités de Turin et de Venise (Italie). La surveillance des sites archéologiques au moyen des images satellitaires polSAR représente un avantage considérable pour la recherche archéologique, alors que les anomalies détectées peuvent concerner les opérations de fouille ou être vérifiées au sol, comme démontré dans ce manuscrit, ou encore elles peuvent contribuer à la réalisation des plans d'intervention pour les sites archéologiques en péril. / Aim of PhD research is to exploit SAR Polarimetry technique for the identification of surface and subsurface archaeological features in the site of Gebel Barkal (Sudan), inscribed in the UNESCO World Heritage List since 2003. Sand penetration capability of both C-band and L-band sensors are discussed analysing archived ALOS PALSAR and RADARSAT-2 specifically acquired (2012-2013) images. Moreover, the research activity illustrates the potential of integrating SAR polarimetric and optical satellite data in a dedicated GIS project, realised in collaboration with the Universities of Turin and Venice (Italy). The monitoring of ancient sites by means of remotely acquired polarimetric SAR data represents a benefit for the archaeological research, where detected anomalies can address archaeological excavations or ground truth verification, as shown in the PhD dissertation, and where threatening factors affect the integrity of a cultural site.

Cette recherche a pour objectif d'étudier l'impact des micro-ondes sur des structures archéologiques toujours visibles de l'ancienne cité de Samarra, la capitale de l'ancien Abbasside Califat situé en Iraq. Trois zones ont été sélectionnées pour cette recherche Doctorale : une cité octogonale, trois champs de courses et la cité de al-Mutawakkiliyya. Les menaces à lesquelles le site est exposé et son importance historique ont permis d'inscrire la cité sur la liste UNESCO des sites en danger (2007). Cela a donné une raison de plus pour enquêter sur cette zone au moyen des capteurs SAR RADARSAT-2 et ALOS PALSAR. Le potentiel de l'imagerie SAR, en fait, est bien connu, en particulier pour les régions du Monde où des enquêtes in situ ne sont pas autorisées en raison de l'instabilité politique (comme dans le cas de Samarra) et en raison de la possibilité d'acquérir quelque soit la couverture nuageuse et dans tout genre d'illumination (jour / nuit). / This work has as goal to study the microwaves behavior over the archaeological structures still visible in the historical city of Samarra, the capital of the ancient Abbasid Caliphate located in Iraq. Three areas were taken into account for the Ph.D. research: an octagonal city, three racecourses stadiums and the city of al-Mutawakkiliyya. Threats to which the site is exposed and its historical importance let the city to be inscribed in the list of UNESCO sites in danger (2007). This gave a reason more to investigate this area by means of SAR RADARSAT-2 and ALOS PALSAR satellites. SAR potentiality, in fact, is well known, in particular for those areas of the World where surveys in situ are not allowed because of political instability (as in the case of Samarra) and because of the possibility of acquiring with any cloud cover conditions and in any kind of illumination (day/night).

Orientador: Gustavo Vitti Môro / Coorientador: Cristiano Zerbato / Coorientador: Antonio Sergio Ferraudo / Banca: Teresa Cristina Tarlé Pissarra / Banca: Viviane Formice Vianna / Resumo: O melhoramento vegetal, além de buscar as características de interesse, busca também otimizar o processo. Sendo assim, quando há correlação entre as características de interesse e uma de fácil avaliação, abre-se a vertente para a seleção indireta. A utilização de sensores na agricultura possibilita a avaliação sem contato físico, podendo ser uma nova ferramenta na seleção indireta, visando otimizar tempo, mão de obra, custo e o processo. Objetivou-se estudar a relação entre o Índice de Vegetação por Diferença Normalizada (NDVI) e os caracteres agronômicos, na seleção indireta em milho e selecionar os genótipos superiores utilizando técnicas multivariadas. O experimento foi conduzido na segunda safra de 2016, sendo realizadas as medições de NDVI via sensor ativo terrestre, a cada 15 dias após a emergência das plântulas e as avaliações agronômicas de campo considerando os caracteres: altura de planta, altura da espiga principal, acamamento, quebramento, estande e produtividade. O conjunto de variáveis obtidas foram submetidas as análises multivariadas de fatores e de componentes principais. A análise de fatores detectou, no primeiro fator, correspondências positivas entre as variáveis, altura de planta, altura de espiga e produtividade, no segundo fator NDVI-80, NDVI-95 e acamamento mais quebramento e no terceiro fator NDVI-15 e estande. Os gráficos biplots gerados pelos componentes principais, juntamente com análise de ganho de seleção permitiram identificar o genótip... (Resumo completo, clicar acesso eletrônico abaixo) / Abstract: The plant breeding look beyond the characteristics of interest, look to optimize process, so when there is a correlation between the characteristics of interest and one of easy evaluation, a strand is opened for indirect selection. The use of sensors in agriculture makes possible to evaluate without physical contact and it can be a new tool in the indirect selection, aiming to optimize time, work, cost and optimize process. The objective of this study was to analyse the relationship between Normalized Difference Vegetation Index (NDVI) and agronomic traits, in the indirect selection and to select superior genotypes of maize by multivariate analyse. The experiment was conducted in the second crop of 2016, and NDVI was measurements by an active sensor every 15 day after seedling and field agronomic traits were evaluated considering the following characteristics: plant height, ear height, stalk lodging, stalk breakage, stand and yield. With these data were processed the factor analyzes, principal component and gain selection. The factor analysis detected positive correspondences between the variables, plant height, ear height and yield with factor 1; NDVI-80, NDVI-95 and stalk lodging plus stalk breakage with factor 2; NDVI- 15 and stand with factor 3. The graphics biplots generated by the principal components with gain selection analyze allowed to identify the best genotype, where we could identify the genotype 3 as the most promising, because it present lower lodgin... (Complete abstract click electronic access below) / Mestre

Grasslands are a significant source of biodiversity in farmed landscapes that is important to monitor. New generation satellites such as Sentinel-2 offer new opportunities for grassland’s monitoring thanks to their combined high spatial and temporal resolutions. Conversely, the new type of data provided by these sensors involves big data and high dimensional issues because of the increasing number of pixels to process and the large number of spectro-temporal variables. This thesis explores the potential of the new generation satellites to monitor biodiversity and factors that influence biodiversity in semi-natural grasslands. Tools suitable for the statistical analysis of grasslands using dense satellite image time series (SITS) with high spatial resolution are provided. First, we show that the spectro-temporal response of grasslands is characterized by its variability within and among the grasslands. Then, for the statistical analysis, grasslands are modeled at the object level to be consistent with ecological models that represent grasslands at the field scale. We propose to model the distribution of pixels in a grassland by a Gaussian distribution. Following this modeling, similarity measures between two Gaussian distributions robust to the high dimension are developed for the lassification of grasslands using dense SITS: the High-Dimensional Kullback-Leibler Divergence and the -Gaussian Mean Kernel. The latter outperforms conventional methods used with Support Vector Machines for the classification of grasslands according to their management practices and to their age. Finally, indicators of grassland biodiversity issued from dense SITS are proposed through spectro-temporal heterogeneity measures derived from the unsupervised clustering of grasslands. Their correlation with the Shannon index is significant but low. The results suggest that the spectro-temporal variations measured from SITS at a spatial resolution of 10 meters covering the period when the practices occur are more related to the intensity of management practices than to the species diversity. Therefore, although the spatial and spectral properties of Sentinel-2 seem limited to assess the species diversity in grasslands directly, this satellite should make possible the continuous monitoring of factors influencing biodiversity in grasslands. In this thesis, we provided methods that account for the heterogeneity within grasslands and enable the use of all the spectral and temporal information provided by new generation satellites.

[pt] A técnica de análise de imagens conhecida pelo acrônimo de GEOBIA (do inglês Geographic Object Based Image Analysis) torna possível a exploração de uma série de novos recursos no processo de classificação de imagens de sensoriamento remoto, em comparação com as alternativas tradicionais baseadas em pixel. Esta possibilidade resulta da introdução de uma etapa de segmentação no processo de análise. Os novos recursos referem-se às propriedades espectrais, texturais, morfológicas e topológicas computadas para os diferentes segmentos de imagem. A abordagem de segmentação habitual encontrada na maioria dos trabalhos de GEOBIA depende de uma hierarquia de segmentações, cada nível de hierarquia associado a um número de classes de objetos caracterizados por tamanhos similares, ou seja, detectáveis em uma determinada escala. A prática usual, porém, não considera segmentações específicas para cada uma das classes de interesse no problema de interpretação, agrupando objetos de mesma escala em um procedimento de segmentação única, ou seja, usando o mesmo algoritmo e parâmetros. A tese investigada neste trabalho baseia-se na suposição de que, se segmentações não são especializadas para cada classe de objeto, então muitos atributos a eles relacionados não podem ser devidamente explorados no processo de classificação. A metodologia proposta baseia-se em uma regra específica para resolver eventuais conflitos espaciais entre as diferentes segmentações. Os resultados experimentais obtidos com base nos experimentos realizados apresentaram um desempenho melhor que o de costume, isto é, produziu melhores resultados de classificação, na maior parte dos problemas de interpretação investigados. / [en] Geographic Object-Based Image Analysis (GEOBIA) makes it possible to exploit a number of new features in the remote sensing image classification process in comparison to the traditional pixel-based alternatives. Such possibility arises from the introduction of a segmentation step in the analysis process. The new features refer to aggregated spectral pixel values, textural, morphological and topological properties computed for the different image segments. The usual segmentation approach found in most GEOBIA works relies on a hierarchy of segmentations, each hierarchy level associated to a number of classes of objects characterized by similar sizes, i.e., which are detectable at a particular scale. The usual practice, therefore, does not consider specific, independent segmentations for each class of interest in the interpretation problem, grouping objects at the same scale through a single segmentation procedure, for instance, using the same algorithm and parameters. The thesis investigated in this work lied on the assumption that if segmentations are not specialized for each object class, then many object features cannot be properly exploited in the classification process. The proposed approach relies on a specific rule to solve eventual spatial conflicts among different segmentations. The experimental results have showed that the proposed approach performed better, i.e., produced better classification results, than the usual one in most of the investigated interpretation problems.

<p>A critical consequence of agriculturally managed lands is the <a></a>transport of nutrients and sediment to fresh water systems, which is ultimately responsible for a range of adverse impacts on human and environmental health. In the U.S. alone, over half of streams and rivers are classified as impaired, with agriculture as the primary contributor. To address deterioration of water quality, there is a need for reliable tools and mathematical models to monitor and predict impacts to water quantity and quality. Soil water content is a key variable in representing environmental systems, linking and driving hydrologic, climate, and biogeochemical cycles; however, the influence of soil water simulations on model predictions is not well characterized, particularly for water quality. Moreover, while soil moisture estimation is the focus of multiple remote sensing missions, defining its potential for use in water quality models remains an open question. The goal of this research is to test whether updating model soil water process representation or model soil water estimates can provide better overall predictive confidence in estimates of both soil moisture and water quality. A widely-used ecohydrologic model, the Soil and Water Assessment Tool (SWAT), was used to evaluate four objectives: 1) investigate the potential of a gridded version of the SWAT model for use with similarly gridded, remote sensing data products, 2) determine the sensitivity of model predictions to changes in soil water content, 3) implement and test a more physically representative soil water percolation algorithm, and 4) perform practical data assimilation experiments using remote sensing data products, focusing on the effects of soil water updates on water quality predictions. With the exception of the first objective, model source code was modified to investigate the relative influence and effect of soil water on overall model predictions. Results suggested that use of the SWAT grid model was currently not viable given practical computational constraints. While the advantages provided by the gridded approach are likely useful for small scale watersheds (< 500 km2), the spatial resolution necessary to run the simulation was too coarse, such that many of the benefits of the gridded approach are negated. Sensitivity tests demonstrated a strong response of model predictions to perturbations in soil moisture. Effects were highly process dependent, where water quality was particularly sensitive to changes in both transport and transformation processes. Model response was reliant upon a default thresholding behavior that restricts subsurface flow and redistribution processes below field capacity. An alternative approach that removed this threshold and keyed processes to relative saturation showed improvement by allowing a more realistic range of soil moisture and a reduction of flushing behavior. This approach was further extended to test against baseline satellite data assimilation experiments; however, did not conclusively outperform the original model simulations. Nevertheless, overall, data assimilation experiments using a remote sensing surface soil moisture data product from the NASA Soil Moisture Active/Passive (SMAP) mission were able to correct for a dry bias in the model simulations and reduce error. Data assimilation updates significantly impacted flow predictions, generally by increasing the dominant contributing flow process. This led to substantial differences between two test sites, where landscape and seasonal characteristics moderated the impact of data assimilation updates to hydrologic, water quality, and crop yield predictions. While the findings illustrate the potential to improve predictions, continued future efforts to refine soil water process representation and optimize data assimilation with longer time series are needed. The dependence of ecohydrologic model predictions on soil moisture highlighted by this research underscores the importance and challenge of effectively representing a complex, physically-based process. As essential decision support systems rely on modeling analyses, improving prediction accuracy is vital.</p>

[pt] O presente trabalho teve como unidade de estudo a bacia hidrográfica do rio Cachoeira, localizada na vertente sul do Maciço da Tijuca, município do Rio de Janeiro. O objetivo geral do estudo foi analisar os efeitos das alterações na cobertura vegetal dessa bacia sobre seu comportamento hidrológico. As componentes hidrológicas selecionadas para análise foram a evapotranspiração real e a vazão média na exutória da bacia. Através do método do balanço hídrico de Thornthwaite e Mather, aplicado de forma sequencial, foi estimada uma série mensal de evapotranspirações reais da bacia. Foram utilizados nesse método de balanço hídrico dados de temperaturas médias do ar e totais precipitados na região. A série de vazões médias na exutória da bacia foi obtida utilizando-se como referência as estações fluviométricas Capela Mayrink e Itanhangá, ambas situadas no interior da bacia. Por meio de tecnologias de geoprocessamento e sensoriamento remoto foram mapeadas dez imagens do satélite Landsat-5/TM, estimando-se assim as alterações ocorridas na cobertura vegetal da bacia. Finalmente buscou-se estabelecer correlações entre as variações da cobertura vegetal e das componentes hidrológicas selecionadas. O resultado obtido para a evapotranspiração real foi satisfatório, indicando uma relação direta com a dinâmica da cobertura da bacia. Entretanto não foi possível estabelecer para a vazão média uma correlação de qualidade semelhante. Através desse estudo adquiriu-se um melhor entendimento sobre a influência da variação da cobertura vegetal no comportamento hidrológico da bacia hidrográfica do rio Cachoeira. / [en] The general objective of the study was to analyze the effects of land cover changes on hydrological processes of the Cachoeira river watershed. The hydrological components selected for analysis were the real evapotranspiration and the mean flow at the exutory of the watershed. Through the Thornthwaite and Mather water balance method, applied sequentially to the period between 1997 and 2010, a monthly series of real evapotranspiration was established. The choice of this method was based on the compatibility of the data required by the method with the hydrometeorological available data. Monthly series of average air temperatures and total precipitations was used to calculate the water balance of Thornthwaite and Mather. The average air temperature data were obtained from the pluviometric station Alto da Boa Vista, located around the watershed. In order to obtain a complete and consistent data series of average air temperatures, a correlation was established of these data with the data from the climatological station of the city of Rio de Janeiro. Through this correlation was possible to confirm the consistency of the series of air temperatures and fill any gaps from 1997 until the year 2010. In this period the monthly series of average air temperatures had a mean of 22,1 Celsius degrees, ranging from a maximum of 27,5 Celsius degrees and a minimum of 17,8 Celsius degrees. The pluviometric data used as reference for the average rainfall over the Cachoeira river watershed, was recorded at the pluviometric station Capela Mayrink, located within the watershed. The consistency of this series has been verified before the other pluviometric stations around the watershed by the method of the double mass, which could confirm the consistency because no deviations were found in the precipitation behavior over time. The average annual precipitation over the watershed was 2.181 mm. The highest estimated values of real evapotranspiration were found in the rainy season, between December and March. The estimated average value of annual real evapotranspiration during the studied period was 1.056 mm, ranging between 1.007 mm and 1166 mm. This average value corresponds to approximately 50 per cent of total annual precipitated, in other words, it is indicated that half of the precipitation over the watershed would return to the atmosphere by the processes of the hydrological cycle. Besides the real evapotranspiration, the water balance also provided an estimation of the water extract of the watershed, calculating components as water deficit, water surplus and soil water storage. The mean monthly flow series at the exutory of the watershed was obtained by referencing the data recorded at the fluviometric stations Capela Mayrink and Itanhangá, both located within the watershed. The data from these stations have gone through a consistency analysis, where their fluviometric levels were verified together and their rating curves were elaborated to represent an adequate adjustment to their liquid discharge measurements. After the consistency analysis, the fluviometric levels data were transformed into flows, through the rating curves. The methodology adopted to generate the flow series at the exutory of the watershed consisted primarily in the extension of the flows series at Itanhangá station through correlation with the flows at Capela Mayrink station. Subsequently, the extended flow series at Itanhangá station was transferred to the exutory location by proportionality between drainage areas. Due to the lack of local data, it was not possible to obtain a mean monthly flow series without gaps. The comparison of the flow data with the precipitation data, obtained for the Cachoeira river watershed, showed a coherent behavior over the years. The annual variation of rainfall in the watershed was accompanied by the flow. It was possible to estimate the changes in land cover during the period from 1988 to 2010 using geoprocessing and remote sensing technologies, available at the extension Spatial Analyst Tool from the software ArcGIS 9.3. In order to obtain this data, a geographic information system was developed for the Cachoeira river watershed, composed by a digital terrain model, obtained from the Shuttle Radar Topography Mission (SRTM), and by ten digital images obtained from the satellite Landsat-5/TM, spaced in average every two years during the studied period. The digital terrain model was used to generate the information grids of Flow direction and Flow accumulation. Through these grids, the watershed and the drainage areas of the fluviometric stations could be automatically delineated. The delimitation of the fluviometric stations drainage areas was done in order to verify the official areas mentioned in the inventory stations of the Brazilian National Water Agency. Although the calculated values did show differences in comparison with the official ones, they were used in the study, considering that the relative errors are minimized when using the same geographic basis. Initially the application of digital processing techniques on satellite images consisted of a combination of bands 5, 4, 3, to form the color composite R, G, B. All images were georeferenced at the same control points in the UTM projection system, using the Datum WGS-84, Zone 23 South. Subsequently the images were classified using the supervised classification maximum likelihood. To characterize the dynamics of land cover over time, two thematic classes were chosen: Forest Area, which has forest cover and others natural features not modified by human activities and Non-Forest Area, which includes urbanized areas and most areas that original feature has been changed as a result of human activities and. The signature samples collected for each training were simple and spatially well distributed, within the region of the studied watershed. Since the supervised classification was an automatic process, the thematic products generated showed errors, identified as isolated cells outside the context of the classes, which left the areas fragmented. In order to work around these errors and provide uniformity of the mapped classes a post-classification process was done on the images by applying a majority filter, which replaces isolated cells based on the majority of their contiguous neighboring cells. Even so, a small portion of the thematic products still showed classification errors, so they were manually edited to become more representative. Ten thematic maps of land cover for the Cachoeira river watershed were generated as products of these processes. The validation of each thematic map classification was verified through the confusion matrix. Considering that only two thematic classes with distinct characteristics were used, the performance of the confusion matrix was tendentious and insufficient to ensure the accuracy of the classification. In order to evaluate the quality of the thematic maps obtained, the thematic map generated for 2010 was compared with the official one, provided by the Municipal Secretariat of Environmental of Rio de Janeiro (SMAC). This comparison could validate the consistency of the thematic map of 2010, believing that the other thematic maps also represent an estimate of the land cover reality from past period. The ten thematic maps could estimate the history of land cover changes on the Cachoeira river watershed. It was observed at the maps that there were changes in the shape of the occupation of the watershed, however, in accordance with the estimated percentages, the evolution of land cover in the watershed had a steady behavior over the years. The consecutive differences found did not exceed the order of 3 per cent. Finally, correlations were established between the variation of the areas with forest coverage and the variation of the selected hydrological components. The result obtained at the real evapotranspiration correlation was satisfactory, which indicates a direct relationship between this hydrological component and the watershed land cover dynamic. However it was not possible to establish a correlation of similar quality with the mean flow. This study could contribute as an exercise to aggregate knowledge about the influence of land cover on hydrological processes over time.

L'étude des sols argileux fait l'objet de nombreux travaux motivés par leur rôle dans les processus d'érosion, les catastrophes naturelles et l'agriculture de précision. La caractérisation du contenu en argiles gonflantes du sol est aussi nécessaire pour évaluer la traficabilité d’une région ou le risque de retrait-gonflement des sols, responsable d’engendrer des dégâts sur le bâti. En effet, les argiles gonflantes sont des smectites qu'il faut différencier des autres types d'argiles telles que l'illite ou la kaolinite, en milieu tempéré. Les techniques traditionnelles pour réaliser la cartographie des minéraux argileux des sols sont en général couteuses, financièrement et en temps et sont basées sur des campagnes terrain intensives simultanément à des acquisitions photographiques afin de spatialiser l'information qui reste qualitative. La télédétection hyperspectrale est une technique potentiellement intéressante pour obtenir des cartes d'argile plus précises et à moindre coût. Néanmoins, elle est limitée par le fait que (i) les minéraux sont mélangés de manière intime dans les sols avec d’autres composants; mais aussi que (ii) à l’échelle aéroportée, le signal réfléchi au sein d’un pixel (résolution spatiale de l’ordre du mètre) comprend de la végétation en plus du sol nu. Ces phénomènes de mélange, aux échelles microscopique et macroscopique, rendent difficile l’estimation du contenu en argiles minéralogiques. Le développement des drones ainsi que leur possibilité d'embarquer de nouvelles caméras hyperspectrales couvrant l'ensemble du spectre [0,4 - 2,5 µm] avec une haute résolution spatiale (environ 10 cm) et un signal à bruit élevé ouvrent la voie à un inventaire plus précis des argiles.L'objectif de cette thèse est de montrer l'intérêt de l'utilisation de méthodes de démélange sur des données hyperspectrales à très haute résolution spatiale pour estimer le contenu en minéraux argileux du sol, et plus précisément des argiles responsables du retrait-gonflement, les smectites. Dans un premier temps, les méthodes existantes de détection, de caractérisation des différents types d'argile et d'estimation de leur abondance sont présentées. Les potentialités des méthodes de démélange existantes dans la littérature pour l’estimation du contenu en minéraux argileux des sols sont mises en avant. Dans un second temps, les méthodes de démélange sont utilisées sur une base de données d’images hyperspectrale acquises en laboratoire de mélanges contrôlés minéraux contenant des argiles (montmorillonite, illite, kaolinite) et d’autres minéraux présents dans les sols (quartz, calcite). Comme les minéraux sont mélangés de manière intime, des méthodes de démélange, linéaires et non-linéaires sont décrites et comparées. Néanmoins, les algorithmes non-linéaires ont des performances similaires aux algorithmes linéaires. De plus, l’effet de la variabilité des données sur la précision de l’estimation des abondances a pu être réduit en utilisant des prétraitements spectraux. Dans une dernière étape, la comparaison des méthodes de démélange sont étendues à des mesures en environnement extérieur. Cette analyse repose sur une campagne de mesure en extérieur réalisée pour la mesure d'images hyperspectrales acquises depuis une nacelle (12 m de hauteur environ, 1,5 cm de résolution spatiale), et l’acquisition d’échantillons prélevés et analysés par DRX (données quantitatives des abondances des minéraux) pour validation. Cette dernière phase permet d'analyser l'impact d'un sol naturel (composé d'un mélange minéralogique, des matières carbonées telles que la cellulose, et ayant une rugosité de surface…) sur les méthodes de démélange. Les performances obtenues (moins de 15% RMSE sur l’estimation de la montmorillonite) permettent d’ouvrir des perspectives quant à l’application de ces méthodes sur des capteurs embarqués par drone, pour la cartographie de la traficabilité et de l’aléa de retrait-gonflement des sols. / Clayey soils are studied because of the importance of soils in erosion processes, natural disasters and precision agriculture. Mapping of clay mineralogy is essential for surveying and predicting trafficability and ground instability hazards, such as shrink-swelling, in order to cope with damages caused by expansive soils on infrastructures. Clay minerals in temperate zone soils are mainly divided in smectites, which highly contribute to soil swelling, illite and kaolinite. Geotechnical engineering practice for clayey soils mapping are expensive and time-consuming. Indeed, it is based on field and extensive laboratory studies. In addition, spatial distribution of clay is assessed using aerial photographs and low-scale geological maps. Thereby, small heterogeneities in geological features are rarely detected, and spatial information remains qualitative. Hyperspectral remote sensing could be an alternative to conventional methods for clay mapping. However, this method is limited by two facts: (i) soils are an intimate mixture of minerals, and (ii) vegetation is mixed with bare soil within airborne sensors pixels (meter range of spatial resolution). Those mixtures (at microscopic and macroscopic scales) mask clays specific spectral signatures and limit clay mineral quantification. Recent development in UAV offers new possibilities for carrying hyperspectral cameras in the reflective domain [0.4 – 2.5 µm], and obtaining data with higher SNR and resolution (10 cm). These advances open new perspectives for accurate and less expensive clay maps. This PhD thesis aims to present the potentiality of clay mapping in soils using very-high spatial resolution hyperspectral data, and more specifically, to estimate swelling clay minerals (smectite) abundances. First, existing methods of detection, and abundance estimation of clay minerals are presented. Second, unmixing methods are used on a database of hyperspectral images of controlled mixtures with different abundances of clay minerals, (illite, montmorillonite and kaolinite) and other minerals existing in soils (quartz, calcite). Due to the intimate nature of mixtures, linear and non-linear unmixing methods are described and compared. However, linear and nonlinear algorithms exhibit similar performances. Moreover, the accuracy of estimation of abundances of mineral clay increased using spectral preprocessings. Regarding the third step, field measurements are used to assess clay unmixing methods. This study is based on an outdoor experiment which acquired hyperspectral images from a bucket truck (12 m elevation, 1.5 cm ground sampling distance), and a sampling collection analyzed by XRD (quantitative analysis of mineral abundances) for validation. This step analyzed effects of a natural soil, (with organic matter, a larger diversity of mineralogical components and with surface roughness) on unmixing methods tested with laboratory data. Obtained performances (less than 15% RMSE for montmorillonite estimation) allow perspectives to apply these methods on data obtained with UAV sensors, for trafficability or expansive soils mapping purposes.

Cette thèse a mis en avant l'intérêt de l'utilisation de données à très haute résolution spatiale et spectrale pour la caractérisation de biodiversité végétale en zone de montagne. D'une part, il a été mis en évidence que des données hyperspectrales (in situ ou aéroportées) permettent de discriminer des habitats végétaux dans une tourbière de montagne. La difficulté de cette étude provient de la forte hétérogénéité qui existe au sein d'une toubière qui connait de forts gradients floristiques et la définition des classes d'habitats qui regroupent plusieurs espèces végétales (parfois communes d'une classe à l'autre). Plus précisément, cette thèse a permis de mettre en évidence que la discrimination pouvait s'effectuer selon trois approches : à partir de mesures de similarité appliquées à la signature spectrale ; en appliquant une classification supervisée qui prend en compte l'information locale (indices spectraux de végétation) ou l'information globale (différents domaines spectraux). Les meilleurs résultats pour distinguer ces différentes classes d'habitats ne sont pas obtenus avec la signature spectrale mais avec des signatures spectrales transformées (CRDR) dans le domaine [350-1350 nm]. Les indices spectraux de végétation qui ont été sélectionnés à partir d'une base non exhaustive, qui caractérise d'autres espèces végétales, sont principalement situés aussi dans ce domaine spectral. De plus, cette thèse a mis en évidence l'intérêt d'appliquer un classifieur peu utilisé pour la classification mais plutôt pour la réduction de dimension (RLR). Une cartographie fine des habitats a également été réalisée en s'appuyant sur des données hyperspectrales aéroportées. / This thesis highlighted the interest of using data with very high spatial and spectral resolution for the characterization of plant biodiversity in mountain areas. On the one hand, it has been shown that (in situ or airborne) hyperspectral data can discriminate plant habitats in a mountain peatbog. The difficulty of this study comes from the strong heterogeneity that exists within a bog that has strong floristic gradients and the definition of habitat classes that group several plant species (sometimes common from one class to another). More specifically, this thesis made it possible to highlight that discrimination could be carried out according to three approaches starting from measures of similarity applied to the spectral signature by applying a supervised classification which takes into account local information (spectral indices of vegetation) or global information (different spectral domains). The best results to distinguish these different habitat classes are not obtained with the spectral signature but with transformed spectral signatures (CRDR) in the spectral range[350-1350 nm]. The spectral vegetation indices that have been selected from a non-exhaustive base, which characterizes other plant species, are also mainly located in this spectral range. Moreover, this thesis highlighted the interest of applying a classifier little used for classification but rather for dimension reduction (RLR). Fine mapping of habitats was also carried out using airborne hyperspectral data.

Thesis submitted in fulfilment of the requirements for the degree of Doctor of Philosophy. School of Animal, Plant and Environmental Science, University of the Witwatersrand, Johannesburg, South Africa. October 2016. / Mining has negative impacts on the environment in many different ways. One method developed to quantify some of these impacts is Landscape Function Analysis (LFA) and this has been accepted by some mining companies and regulators. In brief, LFA aims at quantifying the organization of vegetative and landscape components in a landscape into patches along a transect and quantifying, in a relative manner, three basic processes important to landscape functioning, namely: soil stability or susceptibility to erosion, infiltration or runoff, and nutrient cycling or organic matter decomposition. However, LFA is limited in large heterogeneous environments, such as those around mining operations, due to its localized nature, and the man hours required to collect a representative set of measurements for such large and complex environments. Remote sensing using satellite-acquired data can overcome these limitations by sampling the entire environment in a rapid and objective manner. What is required is a method of connecting these satellite-based measurements to LFA measurements and then being able to extrapolate these measurements across the entire mine surface. The aim of this research was to develop a method to use satellite-based hyperspectral imagery to predict landscape function analysis (LFA) using partial least squares regression (PLSR). This was broken down into three objectives: (1) Collection of the LFA data in the field and validation of the LFA indices against other environmental variables collected at the same time, (2) validation of PLSR models predicting LFA indices and various environmental variables from ground-based spectra, and (3) production of risk maps based on predicting LFA indices and above-ground biomass using PLSR models and Hyperion satellite-based hyperspectral imagery. Although the study was based in grasslands at two mining regions, West Wits and Vaal River, a suitable Hyperion image was only available for Vaal River. A minimum of 374 points were sampled for LFA indices, ground-based spectra, above-ground biomass and soil cores along 2880 m of LFA transect from both mine sites. Soil cores were weighed fresh before sieving with a 2 mm sieve to separate root and stone fractions. The sieved soil fraction was tested for pH, EC, SOM, and for the West Wits samples, organic nitrogen and total extractable inorganic nitrogen. There was one modification to the LFA method where grass patches were collapsed into homogenous units as it was deemed not feasible to sample 180 m transects at grass tuft scales of 10 – 30 cm, but other patch definitions followed the LFA manual (Tongway and Hindley, 2004). Evidence suggested that some of the different patch types, in particular the bare/biological soil crust – bare grass – sparse grass patch types, represented successional stages in a continuum although this was not conclusive. There also was evidence that the presence or absence of cattle play a role in some processes active in these grasslands and erosion is mainly through deflation, rain splash and sheet wash. Generally the environmental variables supported the LFA indices although the nutrient cycling index was representative of above-ground nutrient cycling but not below-ground nutrient cycling. Models derived with PLSR to predict the LFA indices from ground-based spectral measurements were strong at both mine sites (West Wits: LFA stability r2 = 0.63, P < 0.0001; LFA infiltration r2 = 0.75, P < 0.0001; LFA nutrient cycling r2 = 0.73, P < 0.0001; Vaal River: LFA stability r2 = 0.39, P < 0.0001, LFA infiltration r2 = 0.72, P < 0.0001, LFA nutrient cycling r2 = 0.54, P < 0.0001), as were PLSR models predicting above-ground biomass (West Wits above-ground biomass r2 = 0.55, P = 0.0003; Vaal River above-ground biomass r2 = 0.79, P < 0.0001) and soil moisture (West Wits soil moisture r2 = 0.45, P = 0.0017; Vaal River soil moisture r2 = 0.68, P < 0.0001). However, for soil organic matter (r2 = 0.50, P < 0.0001) and EC (r2 = 0.63, P < 0.0001), Vaal River had strong prediction models while West Wits had weak models for these variables (r2 = 0.31, P = 0.019 and r2 = 0.10 and P < 0.18, respectively). For EC, the wide range of soil values at Vaal River in association with gypsum crusts, and low values throughout West Wits explained these model results but for soil organic matter, no clear explanation for these site differences was identified. Patch-based models could accurately discriminate between spectrally well-defined patch types such S. plumosum patches but were less successful with patch types that were spectrally similar such as the bare/biological soil crust – bare grass – sparse grass patch continuum. Clustering similar patch types together before PLSR modelling did improve these patch-based spectral models. To test the method proposed to predict LFA indices from satellite-based hyperspectral imagery, a Hyperion image matching 6 transects at Vaal River was acquired by NASA’s EO-1 satellite and downloaded from the USGS Glovis website. LFA transects were partitioned to match and extract pixel spectra from the Hyperion data cube. Thirty-one spectra were separated into calibration (20) and validation (11) data. PLSR models were derived from the calibration data, tested with validation data to select the optimum model, and then applied to the entire Hyperion data cube to produce prediction maps for five LFA indices and above-ground biomass. The patch area index (PAI) produced particularly strong models (r2 = 0.79, P = 0.0003, n =11) with validation data, whereas the landscape organization index (LOI) produced weak models. It is argued that this difference between these two essentially similar indices is related to the fact that the PAI is a 2-dimensional index and the LOI is a 1-dimensional index. This difference in these two indices allowed the PAI to compensate for some burned pixels on the transects by “seeing” the density pattern of grass tufts and patches whereas the linear nature of the LOI was more susceptible to the changing dimensions of patch structure due to the effects of fire. Although validation models for the three LFA indices of soil stability, infiltration and nutrient cycling were strong (r2 = 0.72, P = 0.004; r2 = 0.66, P = 0.008; r2 = 0.70, P = 0.005, n = 9 respectively), prediction maps were confounded by the presence of fire on some transects. The poor quality of the Hyperion imagery also meant great care had to be taken in the selection of models to avoid poor quality prediction maps. The 31 bands from the VNIR (478 – 885 nm) portion of the Hyperion spectra were generally the best for PLSR modelling and prediction maps, presumably because of better signal-to-noise ratios due to higher energy in the shorter wavelengths. With two satellite-based hyperspectral sensors already operational, namely the US Hyperion and the Chinese HJ-1A HSI, and a number expected to be launched by various space agencies in the next few years, this research presents a method to use the strengths of LFA and hyperspectral imagery to model and predict LFA index values and thereby produce risk maps of large, heterogeneous landscapes such as mining environments. As this research documents a method of partitioning the landscape rather than the pixel spectra into pure endmembers, it makes a valuable contribution to the fields of landscape ecology and hyperspectral remote sensing. / LG2017