Participatory science programs, such as the French Breeding Bird Survey, provide large-scale, spatially extensive biodiversity data at a cost and coverage unattainable by professional monitoring alone. However, these datasets come with inherent statistical challenges: observations are incomplete, spatially heterogeneous, and collected under varying observer effort, while the underlying ecological processes exhibit strong spatial dependence through dispersal, competition, and shared environmental drivers. This thesis develops statistical methods to analyze such partially observed, spatially dependent bird population data, with a focus on two complementary questions: how landscape structure shapes bird communities, and how to rigorously predict local population abundance from incomplete observations.

The first part develops a general theoretical framework for predicting local variation of abundance from partially observed, spatially dependent point processes evolving in a random environment. A kernel-based predictor is constructed as a similarity-weighted average of observed transitions between successive states of the population. Under realistic conditions of spatial mixing and stabilization a law of large numbers and a central limit theorem are established for this predictor. These conditions are verified for several classes of point processes relevant to ecology (log-Gaussian Cox processes, Neyman-Scott processes, Gibbs processes) and for explicit birth, death and similarity functions, including a novel construction to estimate local ecological pressure from incomplete data. The theoretical results are validated through simulation studies and applied to real bird abundance data, where the predictor is shown to capture meaningful trends in local abundance despite the sparsity of citizen-science observations.

The second part of the thesis addresses an applied question in landscape ecology: whether agricultural landscapes should be structured according to a land sparing strategy, separating intensive agriculture from conservation areas, or a land sharing strategy, integrating low-intensity farming and biodiversity within the same landscape. Using 810 sites monitored by the French Breeding Bird Survey between 2001 and 2024, bird counts are modeled through spatio-temporal Bayesian negative binomial regressions, fitted via the Integrated Nested Laplace Approximation (INLA), incorporating a Matérn spatial random field, temporal autocorrelation, climatic and land use covariates. The results show that no single land-use strategy benefits all bird groups: forest specialists and species of conservation concern respond positively to land sparing, while generalist species benefit from the higher habitat heterogeneity associated with land sharing, highlighting the need for context-dependent, group-specific conservation strategies.

Together, these contributions illustrate how combining ecological modeling with tools from spatial statistics and stochastic geometry can extract robust, interpretable information from imperfect biodiversity monitoring data, supporting evidence-based conservation and landscape planning.