Recent literature reflects the substantial progress in combining spatial, temporal and spectral capacities for remote sensing applications. As a result, new issues are arising, such as the need for methodologies that can process simultaneously the different dimensions of satellite information. This paper presents PLS regression extended to three-way data in order to integrate multiwavelengths as variables measured at several dates (time-series) and locations with Sentinel-2 at a regional scale. Considering that the multi-collinearity problem is present in remote sensing time-series to estimate one response variable and that the dataset is multidimensional, a multiway partial least squares (N-PLS) regression approach may be relevant to relate image information to ground variables of interest. N-PLS is an extension of the ordinary PLS regression algorithm where the bilinear model of predictors is replaced by a multilinear model. This paper presents a case study within the context of agriculture, conducted on a time-series of Sentinel-2 images covering regional scale scenes of southern France impacted by the heat wave episode that occurred on 28 June 2019. The model has been developed based on available heat wave impact data for 107 vineyard blocks in the Languedoc-Roussillon region and multispectral time-series predictor data for the period May to August 2019. The results validated the effectiveness of the proposed N-PLS method in estimating yield loss from spectral and temporal attributes. The performance of the model was evaluated by the R² obtained on the prediction set (0.661), and the root mean square of error (RMSE), which was 10.7%. Limitations of the approach when dealing with time-series of large-scale images which represent a source of challenges are discussed; however, the N–PLS regression seems to be a suitable choice for analysing complex multispectral imagery data with different spectral domains and with a clear temporal evolution, such as an extreme weather event.

The potential of image proximal sensing for agricultural applications has been a prolific scientific subject in the recent literature. Its main appeal lies in the sensing of precise information about plant status, which is either harder or impossible to extract from lower-resolution downward-looking image sensors such as satellite or drone imagery. Yet, many theoretical and practical problems arise when dealing with proximal sensing, especially on perennial crops such as vineyards. Indeed, vineyards exhibit challenging physical obstacles and many degrees of variability in their layout. In this paper, we present the design of a mobile camera suited to vineyards and harsh experimental conditions, as well as the results and assessments of 8 years’ worth of studies using that camera. These projects ranged from in-field yield estimation (berry counting) to disease detection, providing new insights on typical viticulture problems that could also be generalized to orchard crops. Different recommendations are then provided using small case studies, such as the difficulties related to framing plots with different structures or the mounting of the sensor on a moving vehicle. While results stress the obvious importance and strong benefits of a thorough experimental design, they also indicate some inescapable pitfalls, illustrating the need for more robust image analysis algorithms and better databases. We believe sharing that experience with the scientific community can only benefit the future development of these innovative approaches.

In this study, a daily grape images dataset derived from a set of 7 monitoring cameras (Vinelapse, IMS Laboratory) is presented. Features describing the joint evolution of grape colour, texture and berry sizes were then extracted using image processing. The relevance of a flexible statistical framework (generalized additive model) to model non-linear trends and identify periods of change in the time series is presented. Results show steady changes can be identified at certain periods, while accounting for model uncertainty. This illustrates the potential of daily grape images time series to compare the phenological behaviour of plants within the same vineyard.

Les maladies du bois de la vigne sont responsables de pertes économiques importantes pour la filière viticole. Ces maladies d'origine fongique se manifestent notamment par une dégradation de la partie boisée du matériel végétal et par l'apparition erratique de symptômes caractéristiques sur la partie foliaire. Cette thèse est dédiée à l'étude de ces maladies (principalement l'esca) à l'aide de deux capteurs imageurs en proxidétection.La question de la détection des symptômes visibles est tout d'abord abordée à l'aide d'un capteur couleur RVB permettant d'acquérir une image par pied de manière automatique ou semi-automatique. La reconnaissance des symptômes est abordée en deux étapes, d'abord en considérant la classification à l'échelle de la feuille puis la détection à l'échelle du pied. La particularité de cette étude est l'inclusion de facteurs confondants dans le problème de classification, tirant partie de l'information de forme des symptômes de l'esca pour les différencier d'autres troubles et maladies. Dans ce but, une comparaison entre approches SIFT et approches transfer learning récentes est alors conduite. Les résultats nous poussent alors à considérer une architecture deep learning simple (RetinaNet) pour la détection des symptômes sur les images, permettant d'estimer un niveau d'atteinte pour chaque pied.Le second capteur utilisé, une caméra hyperspectrale couvrant le spectre de 500 nm à 1300 nm, tente de répondre à une problématique plus expérimentale, à savoir le comportement spectral des pieds atteints par la maladie pouvant déboucher sur une détection précoce des pieds malades mais sans symptômes foliaires. Un protocole expérimental et une base de données de spectres sont alors constitués pour l'occasion. Les méthodes de réduction de la dimensionnalité permettent d'exploiter l'information hyperspectrale voire d'isoler les longueurs d'onde associées à chacune des deux classes. Les données ne permettent cependant pas, pour la plage de longueur d'onde mesurée et dans les conditions d'acquisition terrain, de réaliser une détection précoce de la maladie sur les pieds sans symptômes.Les différences et similarités entre chacune de ces deux applications, en terme de constitution de base de données, d'algorithmes, de difficultés et de potentiel d'application en conditions réelles sont discutées tout au long du manuscrit.