<p>Unmanned aerial vehicles (UAVs) equipped with imaging systems and integrated global navigation satellite system/inertial navigation system (GNSS/INS) are used for a variety of applications. Disaster relief, infrastructure monitoring, precision agriculture, and ecological forestry growth monitoring are among some of the applications that utilize UAV imaging systems. For most applications, accurate 3D spatial information from the UAV imaging system is required. Deriving reliable 3D coordinates is conditioned on accurate geometric calibration. Geometric calibration entails both spatial and temporal calibration. Spatial calibration consists of obtaining accurate internal characteristics of the imaging sensor as well as estimating the mounting parameters between the imaging and the GNSS/INS units. Temporal calibration ensures that there is little to no time delay between the image timestamps and corresponding GNSS/INS position and orientation timestamps. Manual and automated spatial calibration have been successfully accomplished on a variety of platforms and sensors including UAVs equipped with frame and push-broom line cameras. However, manual and automated temporal calibration has not been demonstrated on both frame and line camera systems without the use of ground control points (GCPs). This research focuses on manual and automated spatial and temporal system calibration for UAVs equipped with GNSS/INS frame and line camera systems. For frame cameras, the research introduces two approaches (direct and indirect) to correct for time delay between GNSS/INS recorded event markers and actual time of image exposures. To ensure the best estimates of system parameters without the use of ground control points, an optimal flight configuration for system calibration while estimating time delay is rigorously derived. For line camera systems, this research presents the direct approach to estimate system calibration parameters including time delay during the bundle block adjustment. The optimal flight configuration is also rigorously derived for line camera systems and the bias impact analysis is concluded. This shows that the indirect approach is not a feasible solution for push-broom line cameras onboard UAVs due to the limited ability of line cameras to decouple system parameters and is confirmed with experimental results. Lastly, this research demonstrates that for frame and line camera systems, the direct approach can be fully-automated by incorporating structure from motion (SfM) based tie point features. Methods for feature detection and matching for frame and line camera systems are presented. This research also presents the necessary changes in the bundle adjustment with self-calibration to successfully incorporate a large amount of automatically-derived tie points. For frame cameras, the results show that the direct and indirect approach is capable of estimating and correcting this time delay. When a time delay exists and the direct or indirect approach is applied, horizontal accuracy of 1–3 times the ground sampling distance (GSD) can be achieved without the use of any ground control points (GCPs). For line camera systems, the direct results show that when a time delay exists and spatial and temporal calibration is performed, vertical and horizontal accuracy are approximately that of the ground sample distance (GSD) of the sensor. Furthermore, when a large artificial time delay is introduced for line camera systems, the direct approach still achieves accuracy less than the GSD of the system and performs 2.5-8 times better in the horizontal components and up to 18 times better in the vertical component than when temporal calibration is not performed. Lastly, the results show that automated tie points can be successfully extracted for frame and line camera systems and that those tie point features can be incorporated into a fully-automated bundle adjustment with self-calibration including time delay estimation. The results show that this fully-automated calibration accurately estimates system parameters and demonstrates absolute accuracy similar to that of manually-measured tie/checkpoints without the use of GCPs.</p>

<p> </p> <p>Leaf Area Index (LAI) is commonly defined as the total area of a leaf per unit area of the ground. LAI is an important variable for characterizing plant canopy related to the interception of solar radiation. Direct measurement of LAI by destructive sampling is tedious, time-consuming, and labor-intensive. With the advance of remote sensing, studies have explored multispectral and hyperspectral remote sensing image data and LiDAR point clouds as individual sources to estimate LAI indirectly. This study investigates the estimation of LAI for maize row crops over the growing season based on features derived from high resolution LiDAR and hyperspectral data acquired simultaneously from a UAV platform. Support Vector Regression (SVR) models are developed using cross validation and evaluated relative to the contribution of the multi-modality remote sensing data. The study is based on data acquired for experiments in plant breeding and evaluation of nitrogen management practice trials conducted at the Agronomy Center for Research and Education (ACRE) in 2021 and 2022, respectively. Reference data for the models were collected using a LI-COR® LAI-2200-C Plant Canopy Analyzer. Including both LiDAR and hyperspectral data sources in the SVR model improved the 𝑅_ref^2 (relative to 1:1 comparison line), RMSE and Relative RMSE (rRMSE) values for both the plant breeding and nitrogen management practice experiments, although incremental gains were small overall. More importantly, it was observed that the contributions of the LiDAR vs hyperspectral inputs to the models also varied throughout the growing season. </p>

<p>Integrating unoccupied aerial systems (UAS) into agricultural remote sensing has revolutionized several domains, including crop yield estimation. This research arises from the need to combat citrus greening disease, a major threat to citrus production. Accurately estimating crop yields is crucial for evaluating the effectiveness of treatments and controls for this disease. In response, our study examined the efficacy of phenotypic data extracted from multi-temporal RGB and multispectral UAS images in estimating individual citrus tree yields before harvest and then using this as an indicator to analyze the effectiveness of the treatments and control choice.</p> <p>This study presents machine learning-based regression models for estimating individual citrus tree yields, utilizing the diverse features extracted to provide comprehensive insights into the citrus trees under investigation. Four machine learning algorithms, random forest regression, extreme gradient boosting regression, adaptive boosting, and support vector regression, were employed to build the yield estimation models. The experiment was designed in two phases: single-temporal and multi-temporal modeling.</p>

Cette thèse s'inscrit dans un partenariat entre l'INRA et l'INRIA et dans le champs de l'extraction de connaissances à partir de bases de données spatiales. La problématique porte sur la caractérisation et la simulation de paysages agricoles. Plus précisément, nous nous concentrons sur des lignes qui structurent le paysage agricole, telles que les routes, les fossés d'irrigation et les haies. Notre objectif est de modéliser les haies en raison de leur rôle dans de nombreux processus écologiques et environnementaux. Nous étudions les moyens de caractériser les structures de haies sur deux paysages agricoles contrastés, l'un situé dans le sud-Est de la France (majoritairement composé de vergers) et le second en Bretagne (Ouest de la France, de type bocage). Nous déterminons également si, et dans quelles circonstances, la répartition spatiale des haies est structurée par la position des éléments linéaires plus pérennes du paysage tels que les routes et les fossés et l'échelle de ces structures. La démarche d'extraction de connaissances à partir de base de données (ECBD) mise en place comporte différentes étapes de prétraitement et de fouille de données, alliant des méthodes mathématiques et informatiques. La première partie du travail de thèse se concentre sur la création d'un indice spatial statistique, fondé sur une notion géométrique de voisinage et permettant la caractérisation des structures de haies. Celui-Ci a permis de décrire les structures de haies dans le paysage et les résultats montrent qu'elles dépendent des éléments plus pérennes à courte distance et que le voisinage des haies est uniforme au-Delà de 150 mètres. En outre différentes structures de voisinage ont été mises en évidence selon les principales orientations de haies dans le sud-Est de la France, mais pas en Bretagne. La seconde partie du travail de thèse a exploré l'intérêt du couplage de méthodes de linéarisation avec des méthodes de Markov. Les méthodes de linéarisation ont été introduites avec l'utilisation d'une variante des courbes de Hilbert : les chemins de Hilbert adaptatifs. Les données spatiales linéaires ainsi construites ont ensuite été traitées avec les méthodes de Markov. Ces dernières ont l'avantage de pouvoir servir à la fois pour l'apprentissage sur les données réelles et pour la génération de données, dans le cadre, par exemple, de la simulation d'un paysage. Les résultats montrent que ces méthodes couplées permettant un apprentissage et une génération automatique qui capte des caractéristiques des différents paysages. Les premières simulations sont encourageantes malgré le besoin d'un post-Traitement. Finalement, ce travail de thèse a permis la création d'une méthode d'exploration de données spatiales basée sur différents outils et prenant en charge toutes les étapes de l'ECBD classique, depuis la sélection des données jusqu'à la visualisation des résultats. De plus, la construction de cette méthode est telle qu'elle peut servir à son tour à la génération de données, volet nécessaire pour la simulation de paysage / This thesis is part of a partnership between INRA and INRIA in the field of knowledge extraction from spatial databases. The study focuses on the characterization and simulation of agricultural landscapes. More specifically, we focus on linears that structure the agricultural landscape, such as roads, irrigation ditches and hedgerows. Our goal is to model the spatial distribution of hedgerows because of their role in many ecological and environmental processes. We more specifically study how to characterize the spatial structure of hedgerows in two contrasting agricultural landscapes, one located in south-Eastern France (mainly composed of orchards) and the second in Brittany (western France, \emph{bocage}-Type). We determine if the spatial distribution of hedgerows is structured by the position of the more perennial linear landscape features, such as roads and ditches, or not. In such a case, we also detect the circumstances under which this spatial distribution is structured and the scale of these structures. The implementation of the process of Knowledge Discovery in Databases (KDD) is comprised of different preprocessing steps and data mining algorithms which combine mathematical and computational methods. The first part of the thesis focuses on the creation of a statistical spatial index, based on a geometric neighborhood concept and allowing the characterization of structures of hedgerows. Spatial index allows to describe the structures of hedgerows in the landscape. The results show that hedgerows depend on more permanent linear elements at short distances, and that their neighborhood is uniform beyond 150 meters. In addition different neighborhood structures have been identified depending on the orientation of hedgerows in the South-East of France but not in Brittany. The second part of the thesis explores the potential of coupling linearization methods with Markov methods. The linearization methods are based on the use of alternative Hilbert curves: Hilbert adaptive paths. The linearized spatial data thus constructed were then treated with Markov methods. These methods have the advantage of being able to serve both for the machine learning and for the generation of new data, for example in the context of the simulation of a landscape. The results show that the combination of these methods for learning and automatic generation of hedgerows captures some characteristics of the different study landscapes. The first simulations are encouraging despite the need for post-Processing. Finally, this work has enabled the creation of a spatial data mining method based on different tools that support all stages of a classic KDD, from the selection of data to the visualization of results. Furthermore, this method was constructed in such a way that it can also be used for data generation, a component necessary for the simulation of landscapes

<div>LiDAR units onboard airborne and terrestrial platforms have been established as a proven technology for the acquisition of dense point clouds for a wide range of applications, such as digital building model generation, transportation corridor monitoring, precision agriculture, and infrastructure monitoring. Furthermore, integrating such systems with one or more cameras would allow forward and backward projection between imagery and LiDAR data, thus facilitating several high-level data processing activities such as reliable feature extraction and colorization of point clouds. However, the attainment of the full 3D point positioning potential of such systems is contingent on an accurate calibration of the mobile mapping unit as a whole. </div><div> </div><div> This research aims at proposing a calibration procedure for terrestrial multi-unit LiDAR systems to directly estimate the mounting parameters relating several spinning multi-beam laser scanners to the onboard GNSS/INS unit in order to derive point clouds with high positional accuracy. To ensure the accuracy of the estimated mounting parameters, an optimal configuration of target primitives and drive-runs is determined by analyzing the potential impact of bias in mounting parameters of a LiDAR unit on the resultant point cloud for different orientations of target primitives and different drive-run scenarios. This impact is also verified experimentally by simulating a bias in each mounting parameter separately. Next, the optimal configuration is used within an experimental setup to evaluate the performance of the proposed calibration procedure. Then, this proposed multi-unit LiDAR system calibration strategy is extended for multi-LiDAR multi-camera systems in order to allow a simultaneous estimation of the mounting parameters relating the different laser scanners as well as cameras to the onboard GNSS/INS unit. Such a calibration improves the registration accuracy of point clouds derived from LiDAR data and imagery, along with their accuracy with respect to the ground truth. Finally, in order to qualitatively evaluate the calibration results for a generic mobile mapping system and allow the visualization of point clouds, imagery data, and their registration quality, an interface denoted as Image-LiDAR Interactive Visualization Environment (I-LIVE) is developed. Apart from its visualization functions (such as 3D point cloud manipulation and image display/navigation), I-LIVE mainly serves as a tool for the quality control of GNSS/INS-derived trajectory and LiDAR-camera system calibration. </div><div> </div><div> The proposed multi-sensor system calibration procedures are experimentally evaluated by calibrating several mobile mapping platforms with varying number of LiDAR units and cameras. For all cases, the system calibration is seen to attain accuracies better than the ones expected based on the specifications of the involved hardware components, i.e., the LiDAR units, cameras, and GNSS/INS units.</div>

<div>Leaf Area Index (LAI) is an important variable for both for characterizing plant canopy and as an input to many crop models. It is a dimensionless quantity broadly defined as the total one-sided leaf area per unit ground area, and is estimated over agriculture row crops by both direct and indirect methods. Direct methods, which involve destructive sampling, are laborious and time-consuming, while indirect methods such as remote sensing-based approaches have multiple sources of uncertainty. LiDAR (Light Detection and Ranging) remotely sensed data acquired from manned aircraft and UAVs’ have been investigated to estimate LAI based on physical/geometric features such as canopy gap fraction. High-resolution point cloud data acquired with a laser scanner from any platform, including terrestrial laser scanning and mobile mapping systems, contain random noise and outliers. Therefore, outlier detection in LiDAR data is often useful prior to analysis. Applications in agriculture are particularly challenging, as there is typically no prior knowledge of the statistical distribution of points, description of plant complexity, and local point densities, which are crop dependent. This dissertation first explores the effectiveness of using LiDAR data to estimate LAI for row crop plants at multiple times during the growing season from both a wheeled vehicle and an Unmanned Aerial Vehicle (UAV). Linear and nonlinear regression models are investigated for prediction utilizing statistical and plant structure-based features extracted from the LiDAR point cloud data and ground reference obtained from an in-field plant canopy analyzer and leaf area derived from destructive sampling. LAI estimates obtained from support vector regression (SVR) models with a radial basis function (RBF) kernel developed using the wheel-based LiDAR system and UAVs are promising, based on the value of the coefficient of determination (R2) and root mean squared error (RMSE) of the residuals. </div><div>This dissertation also investigates approaches to minimize the impact of outliers on discrete return LiDAR acquired over crops, and specifically for sorghum and maize breeding experiments, by an unmanned aerial vehicle (UAV) and a wheel-based ground platform. Two methods are explored to detect and remove the outliers from the plant datasets. The first is based on surface fitting to noisy point cloud data based on normal and curvature estimation in a local neighborhood. The second utilizes the deep learning framework PointCleanNet. Both methods are applied to individual plants and field-based datasets. To evaluate the method, an F-score and LAI are calculated both before and after outlier removal for both scenarios. Results indicate that the deep learning method for outlier detection is more robust to changes in point densities, level of noise, and shapes. Also, the predicted LAI was improved for the wheel-based vehicle data based on the R2 value and RMSE of residuals. </div><div>The quality of the extracted features depends on the point density and laser penetration of the canopy. Extracting appropriate features is a critical step to have accurate prediction models. Deep learning frameworks are increasingly being used in remote sensing applications. In the last objective of this study, a feature extraction approach is investigated for encoding LiDAR data acquired by UAV platforms multiple times during the growing season over sorghum and maize plant breeding experiments. LAI estimates obtained with these inputs are used to develop support vector regression (SVR) models using plant canopy analyzer data as the ground reference. Results are compared to models based on estimates from physically-based features and evaluated in terms of the coefficient determination (R2). The effects of experimental conditions, including flying height, sensor characteristics, and crop type, are also investigated relative to the estimates of LAI.</div><div><br></div>

<p dir="ltr">Accurate 3D point clouds are crucial in various fields, and the advancement of software algorithms has facilitated the reconstruction of 3D models from high-quality images. Notably, both single-camera and multi-camera systems have gained popularity in obtaining these images. While single-camera setups offer simplicity and cost-effectiveness, multi-camera systems provide a broader field of view and improved coverage. However, a crucial gap persists, a lack of direct comparison and comprehensive analysis regarding the quality of point clouds acquired from each system. This thesis aims to bridge this gap by evaluating the point cloud quality obtained from both single-camera and multi-camera systems, considering various factors such as lighting conditions, camera settings, and the stability of multi-camera setup in the 3D reconstruction process. Our research also aims to provide insights into how these factors influence the quality and performance of the reconstructed point clouds. By understanding the strengths and limitations of each system, researchers and professionals can make informed decisions when selecting the most suitable 3D imaging approach for their specific applications. To achieve these objectives, we designed and utilized a custom rig with three vertically stacked cameras, each equipped with a fixed camera lens, and maintained uniform lighting conditions. Additionally, we employed a single-camera system with a zoom lens and non uniform lighting conditions. Through noise analysis, our results revealed several crucial findings. The single-camera system exhibited relatively higher noise levels, likely due to non-uniform lighting and the use of a zoom lens. In contrast, the multi-camera system demonstrated lower noise levels, which can be attributed to well-lit conditions and the use of fixed lenses. However, within the multi-camera system, instances of significant instability led to a substantial increase in noise levels in the reconstructed point cloud compared to more stable conditions. Our noise analysis showed the multi-camera system preformed better compared to the single-camera system in terms of noise quality. However, it is crucial to recognize that noise detection also revealed the influence of factors like lighting conditions, camera calibration and camera stability of multi-camera systems on the reconstruction process.</p>