Fore[st]cast

Motivation

‍The main current tasks are about the prediction of tree water deficit (TWD) measured at sparse locations in Switzerland but at regular time intervals. Two models are being developed in parallel: one focuses on temporal forecasts and one on spatial prediction. The output from these models would then be fed to the TreeNet online platform as well as integrated with various drought monitoring platforms in Switzerland.

Proposed Approach / Solution

‍SDSC is involved in the development of predictive models tailored for TWD in the context of sparse spatial locations but with relatively high temporal frequency. The first model is an adaptation of the Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS) model introduced by Challu et al. (2023). This recurrent neural network architecture relies on harmonic bases to represent processes at various temporal scales, similar to Fourier bases in spectral analysis. By leveraging both meteorological and environmental variables, as well as past histories of TWD series, this N-HiTS model can forecast TWD for trees with dendrometers with impressive accuracy. The second model is a long short-term memory (LSTM) neural network tailored for spatial prediction while modeling time series at locations with TWD records. The reason for using a recurrent model for spatial prediction is that trees with dendrometers are only at a few sites which do not cover Switzerland well. Thus only variables computed on a regular grid over Switzerland are used as input features to allow the model to learn reliable relations in time in hopes to then predict at arbitrary locations. These predictions would in turn be processed over particular regions, such as large hydrological catchments, to create maps of TWD used in drought monitoring platforms.

Impact

‍When both models achieve satisfactory performance, both in temporal and spatial prediction of TWD, outputs would be directly usable in forest health monitoring, biodiversity forecasts, and drought monitoring platforms. The latter is notably important for the national system for drought monitoring, forecasting and warning developed by the Swiss Federal Office for the Environment (FOEN) jointly with MeteoSwiss and Swisstopo, and whose first iteration is to be deployed in 2025. Using tree water deficit as an indirect measure of (lack of) water availability in the soil, in addition to more direct indices like predicted precipitation and air temperature, would be crucial for agricultural planning and securing drinking water supplies.