Filter by topics
A deep multi-task learning framework is presented in order to tackle the problem of urban change detection in high and very high resolution satellite images. The proposed scheme is able to couple semantic segmentation with fully convolutional long short-term memory (LSTM) networks. Regarding the fully convolutional LSTM structure, it has been designed by replacing the gating mechanisms with convolutional layers. The goal here is to combine spectral and spatial information, while taking advantage of the temporal relationship among the feature matrices avoiding the computationally expensive task of multiplying high dimensional feature vectors. The fully convolutional LSTM blocks are placed on top of each encoding level of a UNet-like deep architecture, capturing in this way temporal information for all the different resolution levels. In addition, an extra decoding branch is integrated for the semantic segmentation of the available categories, providing the network with fruitful supplementary feature attributes during the training procedure. An ensemble of losses combined in a circular way is also employed for the optimization process. The developed methodology is evaluated on three datasets with different spatial and temporal resolution.
In collaboration with CybeleTech
Meteorological data are needed as input for crop models but their quality's impact on outputs is often neglected. The first part of this PhD was to benchmark spatio-temporal methods for weather data from multiple sources (stations, reanalysis datasets, remote sensing). Inferred data was then plugged in parametric crop models (wheat, corn) for which public datasets were used. Dedicated models were also designed, especially for the problem of soil moisture inference, for which numerous open datasets are available making it suitable for statistical methods.
In the context of the agroecological transition, crop diversification at different levels is seen as a mean to decrease inputs while enhancing yield stability. For wheat, variety mixtures are identified as a promising lever. Diversification is lower than with species mixtures, but variety mixtures have the advantage to be more convenient for farmers. The objective of my PhD is to unravel the functioning of wheat variety mixtures in order to select wheat genotypes for higher-yielding and more stable mixtures. This question is tackled through a combination of field experiments and different modelling approaches. We are developing a dynamic model called WHEAMM (Wheat Model for Mixtures) for the assessment of mixture performance at the plant level. WHEAMM will focus on light competition and its impact on tillering. Calibration will be computed with data from fields experiments currently conducted at Le Moulon with an individual follow-up of plants in mixtures
In collaboration with Estelle Kuhn (INRAE)
Joint model for longitudinal and survival data are widely used in practice to model the effect of a dynamic variable on survival duration (the time between an initial moment and the occurrence of an event). However, few theoretical results have been established for this type of model. In this work, we begin to look at the non-linear mixed effects model for which we want to prove the consistency and the asymptotic normality of the estimators under reasonable assumptions. We will then extend this work to joint survival and mixed effects models. Finally, the objective is to apply this joint model to real data to predict the flowering date of maize, taking into account the dynamics of moth attack on the plant (in collaboration with Estelle Kuhn, Judith Legrand and Elodie Marchadier).