Data-driven models are increasingly used in agriculture to optimize irrigation and promote efficient water use. However, their reliability depends on the stability of the underlying data distributions and data drift, which can be driven by seasonal variability, evolving climatic conditions, or extreme events. In this work, we propose a methodology for constructing confidence intervals around statistical moment estimates to characterize the temporal evolution of variable distributions and identify significant shifts in their underlying dynamics. The approach leverages L-moment statistical theory, with particular emphasis on higher-order L-moments (L-skewness and L-kurtosis), to detect data drift in key environmental variables. We assess the performance of L-moments against classical product moments using real-world data collected between 2022 and 2025 by IoT sensors deployed in 20 orchards across various Spanish regions. Results show that L-moments yield more stable and robust estimates under the varying conditions of these variables and also for small sample sizes, and exhibit greater resilience to outliers. This enables more reliable drift detection without mistaking noise or anomalies for actual distributional changes, which is essential for maintaining robust and adaptive models in agricultural systems. While the application of the proposed methodology is illustrated through a smart farming application, it is readily generalizable to other domains.