A Framework for Modelling Spatio-Temporal Trends in Crop Production Using Generalised Additive Models

Rapid monitoring of crop growth, grain exports, and fire patterns in Ukraine

“…We use remote sensing data and novel analytics capable of accommodating gigabytes of data 7,8 to explore which aspects of the value chain are being affected by war. Specifically, we present trends in crop production over a ten year period using satellite datasets relating to crop growth.…”

“…Inter-annual trends in crop growth can be inferred from decomposition of satellite image time-series into inter-annual, seasonal, and spatial sources of variation 7,9,10 . Inspecting the partial effect of year on cropland Normalised Difference Vegetation Index (NDVI) (Figure 1), which is a well-established proxy for crop growth and has been used to forecast winter wheat yield in Ukraine 11 , shows a decline in 2021 and 2022 from a peak in 2020.…”

“…Further, we can compare NDVI values post-invasion to pre-invasion by adding this factor to our statistical inference framework 7,11 . This comparison revealed cropland NDVI to be similar, and somewhat lower, although insignificant at this stage, (-11 units of NDVI, t = -0.011, p = 0.991) following the invasion of Ukraine on 24 th February 2022.…”

“…The methodology is based on the original GAM approach of Wood et al 9 that was extended to scale to gigabytes of data 8 and accommodate the spatio-temporal dependencies through a penalised regression spline basis set. This model structure was developed for inference of spatio-temporal trends in crop productivity 7 . The following structure was used to derive the partial effect of year (long-term trend) on cropland NDVI in Ukraine:…”