Active learning methods for remote sensing image classification

In this paper, we propose two active learning algorithms for semiautomatic definition of training samples in remote sensing image classification. Based on predefined heuristics, the classifier ranks the unlabeled pixels and automatically chooses those that are considered the most valuable for its improvement. Once the pixels have been selected, the analyst labels them manually and the process is iterated. Starting with a small and nonoptimal training set, the model itself builds the optimal set of samples which minimizes the classification error. We have applied the proposed algorithms to a variety of remote sensing data, including very high resolution and hyperspectral images, using support vector machines. Experimental results confirm the consistency of the methods. The required number of training samples can be reduced to 10% using the methods proposed, reaching the same level of accuracy as larger data sets. A comparison with a state-of-the-art active learning method, margin sampling, is provided, highlighting advantages of the methods proposed. The effect of spatial resolution and separability of the classes on the quality of the selection of pixels is also discussed.

Interpolation-free Coregistration and Phase-Correction of Airborne SAR Interferograms

This letter discusses the detection and correction ofresidual motion errors that appear in airborne synthetic apertureradar (SAR) interferograms due to the lack of precision in the navigationsystem. As it is shown, the effect of this lack of precision istwofold: azimuth registration errors and phase azimuth undulations.Up to now, the correction of the former was carried out byestimating the registration error and interpolating, while the latterwas based on the estimation of the phase azimuth undulations tocompensate the phase of the computed interferogram. In this letter,a new correction method is proposed, which avoids the interpolationstep and corrects at the same time the azimuth phase undulations.Additionally, the spectral diversity technique, used to estimateregistration errors, is critically analyzed. Airborne L-bandrepeat-pass interferometric data of the German Aerospace Center(DLR) experimental airborne SAR is used to validate the method

Comparison of the interobserver variability of 2 different methods of dietary assessment in a geriatric ward: a pilot study.

BACKGROUND: Protein-energy malnutrition is highly prevalent in aged populations. Associated clinical, economic, and social burden is important. A valid screening method that would be robust and precise, but also easy, simple, and rapid to apply, is essential for adequate therapeutic management. OBJECTIVES: To compare the interobserver variability of 2 methods measuring food intake: semiquantitative visual estimations made by nurses versus calorie measurements performed by dieticians on the basis of standardized color digital photographs of servings before and after consumption. DESIGN: Observational monocentric pilot study. SETTING/PARTICIPANTS: A geriatric ward. The meals were randomly chosen from the meal tray. The choice was anonymous with respect to the patients who consumed them. MEASUREMENTS: The test method consisted of the estimation of calorie consumption by dieticians on the basis of standardized color digital photographs of servings before and after consumption. The reference method was based on direct visual estimations of the meals by nurses. Food intake was expressed in the form of a percentage of the serving consumed and calorie intake was then calculated by a dietician based on these percentages. The methods were applied with no previous training of the observers. Analysis of variance was performed to compare their interobserver variability. RESULTS: Of 15 meals consumed and initially examined, 6 were assessed with each method. Servings not consumed at all (0% consumption) or entirely consumed by the patient (100% consumption) were not included in the analysis so as to avoid systematic error. The digital photography method showed higher interobserver variability in calorie intake estimations. The difference between the compared methods was statistically significant (P < .03). CONCLUSIONS: Calorie intake measures for geriatric patients are more concordant when estimated in a semiquantitative way. Digital photography for food intake estimation without previous specific training of dieticians should not be considered as a reference method in geriatric settings, as it shows no advantages in terms of interobserver variability.

Passive knee-extension test to measure hamstring tightness: influence of gravity correction

CONTEXT: A passive knee-extension test has been shown to be a reliable method of assessing hamstring tightness, but this method does not take into account the potential effect of gravity on the tested leg. OBJECTIVE: To compare an original passive knee-extension test with 2 adapted methods including gravity's effect on the lower leg. DESIGN: Repeated measures. SETTING: Laboratory. PARTICIPANTS: 20 young track and field athletes (16.6 ± 1.6 y, 177.6 ± 9.2 cm, 75.9 ± 24.8 kg). INTERVENTION: Each subject was tested in a randomized order with 3 different methods: In the original one (M1), passive knee angle was measured with a standard force of 68.7 N (7 kg) applied proximal to the lateral malleolus. The second (M2) and third (M3) methods took into account the relative lower-leg weight (measured respectively by handheld dynamometer and anthropometrical table) to individualize the force applied to assess passive knee angle. MAIN OUTCOME MEASURES: Passive knee angles measured with video-analysis software. RESULTS: No difference in mean individualized applied force was found between M2 and M3, so the authors assessed passive knee angle only with M2. The mean knee angle was different between M1 and M2 (68.8 ± 12.4 vs 73.1 ± 10.6, P < .001). Knee angles in M1 and M2 were correlated (r = .93, P < .001). CONCLUSIONS: Differences in knee angle were found between the original passive knee-extension test and a method with gravity correction. M2 is an improved version of the original method (M1) since it minimizes the effect of gravity. Therefore, we recommend using it rather than M1.

A limited area model intercomparison on the "Montserrat-2000" flash-flood event using statistical and deterministic methods

In the scope of the European project Hydroptimet, INTERREG IIIB-MEDOCC programme, limited area model (LAM) intercomparison of intense events that produced many damages to people and territory is performed. As the comparison is limited to single case studies, the work is not meant to provide a measure of the different models' skill, but to identify the key model factors useful to give a good forecast on such a kind of meteorological phenomena. This work focuses on the Spanish flash-flood event, also known as "Montserrat-2000" event. The study is performed using forecast data from seven operational LAMs, placed at partners' disposal via the Hydroptimet ftp site, and observed data from Catalonia rain gauge network. To improve the event analysis, satellite rainfall estimates have been also considered. For statistical evaluation of quantitative precipitation forecasts (QPFs), several non-parametric skill scores based on contingency tables have been used. Furthermore, for each model run it has been possible to identify Catalonia regions affected by misses and false alarms using contingency table elements. Moreover, the standard "eyeball" analysis of forecast and observed precipitation fields has been supported by the use of a state-of-the-art diagnostic method, the contiguous rain area (CRA) analysis. This method allows to quantify the spatial shift forecast error and to identify the error sources that affected each model forecasts. High-resolution modelling and domain size seem to have a key role for providing a skillful forecast. Further work is needed to support this statement, including verification using a wider observational data set.

Enhanced radar precipitation estimates using a combined clutter and beam blockage correction technique

Weather radar observations are currently the most reliable method for remote sensing of precipitation. However, a number of factors affect the quality of radar observations and may limit seriously automated quantitative applications of radar precipitation estimates such as those required in Numerical Weather Prediction (NWP) data assimilation or in hydrological models. In this paper, a technique to correct two different problems typically present in radar data is presented and evaluated. The aspects dealt with are non-precipitating echoes - caused either by permanent ground clutter or by anomalous propagation of the radar beam (anaprop echoes) - and also topographical beam blockage. The correction technique is based in the computation of realistic beam propagation trajectories based upon recent radiosonde observations instead of assuming standard radio propagation conditions. The correction consists of three different steps: 1) calculation of a Dynamic Elevation Map which provides the minimum clutter-free antenna elevation for each pixel within the radar coverage; 2) correction for residual anaprop, checking the vertical reflectivity gradients within the radar volume; and 3) topographical beam blockage estimation and correction using a geometric optics approach. The technique is evaluated with four case studies in the region of the Po Valley (N Italy) using a C-band Doppler radar and a network of raingauges providing hourly precipitation measurements. The case studies cover different seasons, different radio propagation conditions and also stratiform and convective precipitation type events. After applying the proposed correction, a comparison of the radar precipitation estimates with raingauges indicates a general reduction in both the root mean squared error and the fractional error variance indicating the efficiency and robustness of the procedure. Moreover, the technique presented is not computationally expensive so it seems well suited to be implemented in an operational environment.

Correction for heart rate variability during 3D whole heart MR coronary angiography.

PURPOSE: To evaluate the effect of a real-time adaptive trigger delay on image quality to correct for heart rate variability in 3D whole-heart coronary MR angiography (MRA). MATERIALS AND METHODS: Twelve healthy adults underwent 3D whole-heart coronary MRA with and without the use of an adaptive trigger delay. The moment of minimal coronary artery motion was visually determined on a high temporal resolution MRI. Throughout the scan performed without adaptive trigger delay, trigger delay was kept constant, whereas during the scan performed with adaptive trigger delay, trigger delay was continuously updated after each RR-interval using physiological modeling. Signal-to-noise, contrast-to-noise, vessel length, vessel sharpness, and subjective image quality were compared in a blinded manner. RESULTS: Vessel sharpness improved significantly for the middle segment of the right coronary artery (RCA) with the use of the adaptive trigger delay (52.3 +/- 7.1% versus 48.9 +/- 7.9%, P = 0.026). Subjective image quality was significantly better in the middle segments of the RCA and left anterior descending artery (LAD) when the scan was performed with adaptive trigger delay compared to constant trigger delay. CONCLUSION: Our results demonstrate that the use of an adaptive trigger delay to correct for heart rate variability improves image quality mainly in the middle segments of the RCA and LAD.

Gas identification with tin oxide sensor array and self-organizing maps: adaptive correction of sensor drifts

Low-cost tin oxide gas sensors are inherently nonspecific. In addition, they have several undesirable characteristics such as slow response, nonlinearities, and long-term drifts. This paper shows that the combination of a gas-sensor array together with self-organizing maps (SOM's) permit success in gas classification problems. The system is able to determine the gas present in an atmosphere with error rates lower than 3%. Correction of the sensor's drift with an adaptive SOM has also been investigated

Submillimeter three-dimensional coronary MR angiography with real-time navigator correction: comparison of navigator locations.

Three-dimensional free-breathing coronary magnetic resonance angiography was performed in eight healthy volunteers with use of real-time navigator technology. Images acquired with the navigator localized at the right hemidiaphragm and at the left ventricle were objectively compared. The diaphragmatic navigator was found to be superior for vessel delineation of middle to distal portions of the coronary arteries.

Local and global error models to improve uncertainty quantification

In groundwater applications, Monte Carlo methods are employed to model the uncertainty on geological parameters. However, their brute-force application becomes computationally prohibitive for highly detailed geological descriptions, complex physical processes, and a large number of realizations. The Distance Kernel Method (DKM) overcomes this issue by clustering the realizations in a multidimensional space based on the flow responses obtained by means of an approximate (computationally cheaper) model; then, the uncertainty is estimated from the exact responses that are computed only for one representative realization per cluster (the medoid). Usually, DKM is employed to decrease the size of the sample of realizations that are considered to estimate the uncertainty. We propose to use the information from the approximate responses for uncertainty quantification. The subset of exact solutions provided by DKM is then employed to construct an error model and correct the potential bias of the approximate model. Two error models are devised that both employ the difference between approximate and exact medoid solutions, but differ in the way medoid errors are interpolated to correct the whole set of realizations. The Local Error Model rests upon the clustering defined by DKM and can be seen as a natural way to account for intra-cluster variability; the Global Error Model employs a linear interpolation of all medoid errors regardless of the cluster to which the single realization belongs. These error models are evaluated for an idealized pollution problem in which the uncertainty of the breakthrough curve needs to be estimated. For this numerical test case, we demonstrate that the error models improve the uncertainty quantification provided by the DKM algorithm and are effective in correcting the bias of the estimate computed solely from the MsFV results. The framework presented here is not specific to the methods considered and can be applied to other combinations of approximate models and techniques to select a subset of realizations

Cross validation of experts versus registration methods for target localization in deep brain stimulation.

In the last five years, Deep Brain Stimulation (DBS) has become the most popular and effective surgical technique for the treatent of Parkinson's disease (PD). The Subthalamic Nucleus (STN) is the usual target involved when applying DBS. Unfortunately, the STN is in general not visible in common medical imaging modalities. Therefore, atlas-based segmentation is commonly considered to locate it in the images. In this paper, we propose a scheme that allows both, to perform a comparison between different registration algorithms and to evaluate their ability to locate the STN automatically. Using this scheme we can evaluate the expert variability against the error of the algorithms and we demonstrate that automatic STN location is possible and as accurate as the methods currently used.

New methods to quantify NH3 volatilization from fertilized surface soil with urea

Gaseous N losses from soil are considerable, resulting mostly from ammonia volatilization linked to agricultural activities such as pasture fertilization. The use of simple and accessible measurement methods of such losses is fundamental in the evaluation of the N cycle in agricultural systems. The purpose of this study was to evaluate quantification methods of NH3 volatilization from fertilized surface soil with urea, with minimal influence on the volatilization processes. The greenhouse experiment was arranged in a completely randomized design with 13 treatments and five replications, with the following treatments: (1) Polyurethane foam (density 20 kg m-3) with phosphoric acid solution absorber (foam absorber), installed 1, 5, 10 and 20 cm above the soil surface; (2) Paper filter with sulfuric acid solution absorber (paper absorber, 1, 5, 10 and 20 cm above the soil surface); (3) Sulfuric acid solution absorber (1, 5 and 10 cm above the soil surface); (4) Semi-open static collector; (5) 15N balance (control). The foam absorber placed 1 cm above the soil surface estimated the real daily rate of loss and accumulated loss of NH3N and proved efficient in capturing NH3 volatized from urea-treated soil. The estimates based on acid absorbers 1, 5 and 10 cm above the soil surface and paper absorbers 1 and 5 cm above the soil surface were only realistic for accumulated N-NH3 losses. Foam absorbers can be indicated to quantify accumulated and daily rates of NH3 volatilization losses similarly to an open static chamber, making calibration equations or correction factors unnecessary.

Correction of through-plane deformation artifacts in stimulated echo acquisition mode cardiac imaging.

Attempts to use a stimulated echo acquisition mode (STEAM) in cardiac imaging are impeded by imaging artifacts that result in signal attenuation and nulling of the cardiac tissue. In this work, we present a method to reduce this artifact by acquiring two sets of stimulated echo images with two different demodulations. The resulting two images are combined to recover the signal loss and weighted to compensate for possible deformation-dependent intensity variation. Numerical simulations were used to validate the theory. Also, the proposed correction method was applied to in vivo imaging of normal volunteers (n = 6) and animal models with induced infarction (n = 3). The results show the ability of the method to recover the lost myocardial signal and generate artifact-free black-blood cardiac images.

Real-time motion correction in navigator-gated free-breathing double-oblique submillimeter 3D right coronary artery magnetic resonance angiography.

RATIONALE AND OBJECTIVES: The purpose of this study was the investigation of the impact of real-time adaptive motion correction on image quality in navigator-gated, free-breathing, double-oblique three-dimensional (3D) submillimeter right coronary magnetic resonance angiography (MRA). MATERIALS AND METHODS: Free-breathing 3D right coronary MRA with real-time navigator technology was performed in 10 healthy adult subjects with an in-plane spatial resolution of 700 x 700 microm. Identical double-oblique coronary MR-angiograms were performed with navigator gating alone and combined navigator gating and real-time adaptive motion correction. Quantitative objective parameters of contrast-to-noise ratio (CNR) and vessel sharpness and subjective image quality scores were compared. RESULTS: Superior vessel sharpness, increased CNR, and superior image quality scores were found with combined navigator gating and real-time adaptive motion correction (vs. navigator gating alone; P < 0.01 for all comparisons). CONCLUSION: Real-time adaptive motion correction objectively and subjectively improves image quality in 3D navigator-gated free-breathing double-oblique submillimeter right coronary MRA.

Reliability of methods used for the risk assessment of mixture of micropollutants in aquatic enviromments : from theoretical concepts to field observations

Ces dernières années, de nombreuses recherches ont mis en évidence les effets toxiques des micropolluants organiques pour les espèces de nos lacs et rivières. Cependant, la plupart de ces études se sont focalisées sur la toxicité des substances individuelles, alors que les organismes sont exposés tous les jours à des milliers de substances en mélange. Or les effets de ces cocktails ne sont pas négligeables. Cette thèse de doctorat s'est ainsi intéressée aux modèles permettant de prédire le risque environnemental de ces cocktails pour le milieu aquatique. Le principal objectif a été d'évaluer le risque écologique des mélanges de substances chimiques mesurées dans le Léman, mais aussi d'apporter un regard critique sur les méthodologies utilisées afin de proposer certaines adaptations pour une meilleure estimation du risque. Dans la première partie de ce travail, le risque des mélanges de pesticides et médicaments pour le Rhône et pour le Léman a été établi en utilisant des approches envisagées notamment dans la législation européenne. Il s'agit d'approches de « screening », c'est-à-dire permettant une évaluation générale du risque des mélanges. Une telle approche permet de mettre en évidence les substances les plus problématiques, c'est-à-dire contribuant le plus à la toxicité du mélange. Dans notre cas, il s'agit essentiellement de 4 pesticides. L'étude met également en évidence que toutes les substances, même en trace infime, contribuent à l'effet du mélange. Cette constatation a des implications en terme de gestion de l'environnement. En effet, ceci implique qu'il faut réduire toutes les sources de polluants, et pas seulement les plus problématiques. Mais l'approche proposée présente également un biais important au niveau conceptuel, ce qui rend son utilisation discutable, en dehors d'un screening, et nécessiterait une adaptation au niveau des facteurs de sécurité employés. Dans une deuxième partie, l'étude s'est portée sur l'utilisation des modèles de mélanges dans le calcul de risque environnemental. En effet, les modèles de mélanges ont été développés et validés espèce par espèce, et non pour une évaluation sur l'écosystème en entier. Leur utilisation devrait donc passer par un calcul par espèce, ce qui est rarement fait dû au manque de données écotoxicologiques à disposition. Le but a été donc de comparer, avec des valeurs générées aléatoirement, le calcul de risque effectué selon une méthode rigoureuse, espèce par espèce, avec celui effectué classiquement où les modèles sont appliqués sur l'ensemble de la communauté sans tenir compte des variations inter-espèces. Les résultats sont dans la majorité des cas similaires, ce qui valide l'approche utilisée traditionnellement. En revanche, ce travail a permis de déterminer certains cas où l'application classique peut conduire à une sous- ou sur-estimation du risque. Enfin, une dernière partie de cette thèse s'est intéressée à l'influence que les cocktails de micropolluants ont pu avoir sur les communautés in situ. Pour ce faire, une approche en deux temps a été adoptée. Tout d'abord la toxicité de quatorze herbicides détectés dans le Léman a été déterminée. Sur la période étudiée, de 2004 à 2009, cette toxicité due aux herbicides a diminué, passant de 4% d'espèces affectées à moins de 1%. Ensuite, la question était de savoir si cette diminution de toxicité avait un impact sur le développement de certaines espèces au sein de la communauté des algues. Pour ce faire, l'utilisation statistique a permis d'isoler d'autres facteurs pouvant avoir une influence sur la flore, comme la température de l'eau ou la présence de phosphates, et ainsi de constater quelles espèces se sont révélées avoir été influencées, positivement ou négativement, par la diminution de la toxicité dans le lac au fil du temps. Fait intéressant, une partie d'entre-elles avait déjà montré des comportements similaires dans des études en mésocosmes. En conclusion, ce travail montre qu'il existe des modèles robustes pour prédire le risque des mélanges de micropolluants sur les espèces aquatiques, et qu'ils peuvent être utilisés pour expliquer le rôle des substances dans le fonctionnement des écosystèmes. Toutefois, ces modèles ont bien sûr des limites et des hypothèses sous-jacentes qu'il est important de considérer lors de leur application. - Depuis plusieurs années, les risques que posent les micropolluants organiques pour le milieu aquatique préoccupent grandement les scientifiques ainsi que notre société. En effet, de nombreuses recherches ont mis en évidence les effets toxiques que peuvent avoir ces substances chimiques sur les espèces de nos lacs et rivières, quand elles se retrouvent exposées à des concentrations aiguës ou chroniques. Cependant, la plupart de ces études se sont focalisées sur la toxicité des substances individuelles, c'est à dire considérées séparément. Actuellement, il en est de même dans les procédures de régulation européennes, concernant la partie évaluation du risque pour l'environnement d'une substance. Or, les organismes sont exposés tous les jours à des milliers de substances en mélange, et les effets de ces "cocktails" ne sont pas négligeables. L'évaluation du risque écologique que pose ces mélanges de substances doit donc être abordé par de la manière la plus appropriée et la plus fiable possible. Dans la première partie de cette thèse, nous nous sommes intéressés aux méthodes actuellement envisagées à être intégrées dans les législations européennes pour l'évaluation du risque des mélanges pour le milieu aquatique. Ces méthodes sont basées sur le modèle d'addition des concentrations, avec l'utilisation des valeurs de concentrations des substances estimées sans effet dans le milieu (PNEC), ou à partir des valeurs des concentrations d'effet (CE50) sur certaines espèces d'un niveau trophique avec la prise en compte de facteurs de sécurité. Nous avons appliqué ces méthodes à deux cas spécifiques, le lac Léman et le Rhône situés en Suisse, et discuté les résultats de ces applications. Ces premières étapes d'évaluation ont montré que le risque des mélanges pour ces cas d'étude atteint rapidement une valeur au dessus d'un seuil critique. Cette valeur atteinte est généralement due à deux ou trois substances principales. Les procédures proposées permettent donc d'identifier les substances les plus problématiques pour lesquelles des mesures de gestion, telles que la réduction de leur entrée dans le milieu aquatique, devraient être envisagées. Cependant, nous avons également constaté que le niveau de risque associé à ces mélanges de substances n'est pas négligeable, même sans tenir compte de ces substances principales. En effet, l'accumulation des substances, même en traces infimes, atteint un seuil critique, ce qui devient plus difficile en terme de gestion du risque. En outre, nous avons souligné un manque de fiabilité dans ces procédures, qui peuvent conduire à des résultats contradictoires en terme de risque. Ceci est lié à l'incompatibilité des facteurs de sécurité utilisés dans les différentes méthodes. Dans la deuxième partie de la thèse, nous avons étudié la fiabilité de méthodes plus avancées dans la prédiction de l'effet des mélanges pour les communautés évoluant dans le système aquatique. Ces méthodes reposent sur le modèle d'addition des concentrations (CA) ou d'addition des réponses (RA) appliqués sur les courbes de distribution de la sensibilité des espèces (SSD) aux substances. En effet, les modèles de mélanges ont été développés et validés pour être appliqués espèce par espèce, et non pas sur plusieurs espèces agrégées simultanément dans les courbes SSD. Nous avons ainsi proposé une procédure plus rigoureuse, pour l'évaluation du risque d'un mélange, qui serait d'appliquer d'abord les modèles CA ou RA à chaque espèce séparément, et, dans une deuxième étape, combiner les résultats afin d'établir une courbe SSD du mélange. Malheureusement, cette méthode n'est pas applicable dans la plupart des cas, car elle nécessite trop de données généralement indisponibles. Par conséquent, nous avons comparé, avec des valeurs générées aléatoirement, le calcul de risque effectué selon cette méthode plus rigoureuse, avec celle effectuée traditionnellement, afin de caractériser la robustesse de cette approche qui consiste à appliquer les modèles de mélange sur les courbes SSD. Nos résultats ont montré que l'utilisation de CA directement sur les SSDs peut conduire à une sous-estimation de la concentration du mélange affectant 5 % ou 50% des espèces, en particulier lorsque les substances présentent un grand écart- type dans leur distribution de la sensibilité des espèces. L'application du modèle RA peut quant à lui conduire à une sur- ou sous-estimations, principalement en fonction de la pente des courbes dose- réponse de chaque espèce composant les SSDs. La sous-estimation avec RA devient potentiellement importante lorsque le rapport entre la EC50 et la EC10 de la courbe dose-réponse des espèces est plus petit que 100. Toutefois, la plupart des substances, selon des cas réels, présentent des données d' écotoxicité qui font que le risque du mélange calculé par la méthode des modèles appliqués directement sur les SSDs reste cohérent et surestimerait plutôt légèrement le risque. Ces résultats valident ainsi l'approche utilisée traditionnellement. Néanmoins, il faut garder à l'esprit cette source d'erreur lorsqu'on procède à une évaluation du risque d'un mélange avec cette méthode traditionnelle, en particulier quand les SSD présentent une distribution des données en dehors des limites déterminées dans cette étude. Enfin, dans la dernière partie de cette thèse, nous avons confronté des prédictions de l'effet de mélange avec des changements biologiques observés dans l'environnement. Dans cette étude, nous avons utilisé des données venant d'un suivi à long terme d'un grand lac européen, le lac Léman, ce qui offrait la possibilité d'évaluer dans quelle mesure la prédiction de la toxicité des mélanges d'herbicide expliquait les changements dans la composition de la communauté phytoplanctonique. Ceci à côté d'autres paramètres classiques de limnologie tels que les nutriments. Pour atteindre cet objectif, nous avons déterminé la toxicité des mélanges sur plusieurs années de 14 herbicides régulièrement détectés dans le lac, en utilisant les modèles CA et RA avec les courbes de distribution de la sensibilité des espèces. Un gradient temporel de toxicité décroissant a pu être constaté de 2004 à 2009. Une analyse de redondance et de redondance partielle, a montré que ce gradient explique une partie significative de la variation de la composition de la communauté phytoplanctonique, même après avoir enlevé l'effet de toutes les autres co-variables. De plus, certaines espèces révélées pour avoir été influencées, positivement ou négativement, par la diminution de la toxicité dans le lac au fil du temps, ont montré des comportements similaires dans des études en mésocosmes. On peut en conclure que la toxicité du mélange herbicide est l'un des paramètres clés pour expliquer les changements de phytoplancton dans le lac Léman. En conclusion, il existe diverses méthodes pour prédire le risque des mélanges de micropolluants sur les espèces aquatiques et celui-ci peut jouer un rôle dans le fonctionnement des écosystèmes. Toutefois, ces modèles ont bien sûr des limites et des hypothèses sous-jacentes qu'il est important de considérer lors de leur application, avant d'utiliser leurs résultats pour la gestion des risques environnementaux. - For several years now, the scientists as well as the society is concerned by the aquatic risk organic micropollutants may pose. Indeed, several researches have shown the toxic effects these substances may induce on organisms living in our lakes or rivers, especially when they are exposed to acute or chronic concentrations. However, most of the studies focused on the toxicity of single compounds, i.e. considered individually. The same also goes in the current European regulations concerning the risk assessment procedures for the environment of these substances. But aquatic organisms are typically exposed every day simultaneously to thousands of organic compounds. The toxic effects resulting of these "cocktails" cannot be neglected. The ecological risk assessment of mixtures of such compounds has therefore to be addressed by scientists in the most reliable and appropriate way. In the first part of this thesis, the procedures currently envisioned for the aquatic mixture risk assessment in European legislations are described. These methodologies are based on the mixture model of concentration addition and the use of the predicted no effect concentrations (PNEC) or effect concentrations (EC50) with assessment factors. These principal approaches were applied to two specific case studies, Lake Geneva and the River Rhône in Switzerland, including a discussion of the outcomes of such applications. These first level assessments showed that the mixture risks for these studied cases exceeded rapidly the critical value. This exceeding is generally due to two or three main substances. The proposed procedures allow therefore the identification of the most problematic substances for which management measures, such as a reduction of the entrance to the aquatic environment, should be envisioned. However, it was also showed that the risk levels associated with mixtures of compounds are not negligible, even without considering these main substances. Indeed, it is the sum of the substances that is problematic, which is more challenging in term of risk management. Moreover, a lack of reliability in the procedures was highlighted, which can lead to contradictory results in terms of risk. This result is linked to the inconsistency in the assessment factors applied in the different methods. In the second part of the thesis, the reliability of the more advanced procedures to predict the mixture effect to communities in the aquatic system were investigated. These established methodologies combine the model of concentration addition (CA) or response addition (RA) with species sensitivity distribution curves (SSD). Indeed, the mixture effect predictions were shown to be consistent only when the mixture models are applied on a single species, and not on several species simultaneously aggregated to SSDs. Hence, A more stringent procedure for mixture risk assessment is proposed, that would be to apply first the CA or RA models to each species separately and, in a second step, to combine the results to build an SSD for a mixture. Unfortunately, this methodology is not applicable in most cases, because it requires large data sets usually not available. Therefore, the differences between the two methodologies were studied with datasets created artificially to characterize the robustness of the traditional approach applying models on species sensitivity distribution. The results showed that the use of CA on SSD directly might lead to underestimations of the mixture concentration affecting 5% or 50% of species, especially when substances present a large standard deviation of the distribution from the sensitivity of the species. The application of RA can lead to over- or underestimates, depending mainly on the slope of the dose-response curves of the individual species. The potential underestimation with RA becomes important when the ratio between the EC50 and the EC10 for the dose-response curve of the species composing the SSD are smaller than 100. However, considering common real cases of ecotoxicity data for substances, the mixture risk calculated by the methodology applying mixture models directly on SSDs remains consistent and would rather slightly overestimate the risk. These results can be used as a theoretical validation of the currently applied methodology. Nevertheless, when assessing the risk of mixtures, one has to keep in mind this source of error with this classical methodology, especially when SSDs present a distribution of the data outside the range determined in this study Finally, in the last part of this thesis, we confronted the mixture effect predictions with biological changes observed in the environment. In this study, long-term monitoring of a European great lake, Lake Geneva, provides the opportunity to assess to what extent the predicted toxicity of herbicide mixtures explains the changes in the composition of the phytoplankton community next to other classical limnology parameters such as nutrients. To reach this goal, the gradient of the mixture toxicity of 14 herbicides regularly detected in the lake was calculated, using concentration addition and response addition models. A decreasing temporal gradient of toxicity was observed from 2004 to 2009. Redundancy analysis and partial redundancy analysis showed that this gradient explains a significant portion of the variation in phytoplankton community composition, even when having removed the effect of all other co-variables. Moreover, some species that were revealed to be influenced positively or negatively, by the decrease of toxicity in the lake over time, showed similar behaviors in mesocosms studies. It could be concluded that the herbicide mixture toxicity is one of the key parameters to explain phytoplankton changes in Lake Geneva. To conclude, different methods exist to predict the risk of mixture in the ecosystems. But their reliability varies depending on the underlying hypotheses. One should therefore carefully consider these hypotheses, as well as the limits of the approaches, before using the results for environmental risk management