|Title||CST, a freeware for predicting crop yield from remote sensing or crop model indicators: Illustration with RSA and Ethiopia|
|Author(s)||Kerdiles, H.; Rembold, F.; Leo, O.; Boogaard, H.; Hoek, S.|
|Source||In: 2017 6th International Conference on Agro-Geoinformatics, Agro-Geoinformatics 2017. - Institute of Electrical and Electronics Engineers Inc. - ISBN 9781538638842|
|Event||6th International Conference on Agro-Geoinformatics, Agro-Geoinformatics 2017, Fairfax, 2017-08-07/2017-08-10|
|Department(s)||Earth Observation and Environmental Informatics|
|Publication type||Contribution in proceedings|
|Keyword(s)||crop yield prediction - food security - remote sensing|
CST (Crop Statistics Tool) is a standalone freeware for predicting crop yield statistics using indicators derived from crop models, weather or remote sensing data. The principle of CST is that years similar to the target year (e.g. the current year) should have similar yields, or similar yield deviations from a technological time trend. In practice, CST guides the crop analyst through standard steps: After data screening to identify possible outliers and analysis of time trend, the crop analyst has the choice between the following two approaches to forecast yield: (1) multiple regression analysis in which a linear relationship is calibrated between historical yield data and yield indicators, while accounting for a time trend if present; (2) scenario analysis, whereby CST looks for the years most similar (according to the indicators) to the current year to estimate a yield deviation from the time trend or the average yield. CST allows to assess models with standard statistics and tests as well as warnings, which is especially useful when many indicators are available. Moreover, thanks to batch processing, the crop analyst can test a given model for various dekads, regions or crops. This paper illustrates the interest of CST with two case studies made over Africa and based on the regression approach between crop yields and NDVI or cumulated rainfall at a given dekad. In the first one, South African maize yields at province level over 1987-2015 were found to be well correlated with Vegetation/ProbaV NDVI or CHIRPS rainfall for two of the three main maize producing provinces; for each province, we tested indicators from the 15 dekads between January and May. In the second study, we regressed the 1999-2014 maize yields from the main 26 crop production zones against also NDVI and cumulated rainfall with two different start dates (April and June); we tested 3000 models (26 zones, 15 dekads, 3 single indicators without and with time trend, all indicators together, and finally trend alone) and obtained mixed results: A strong dominance of the time trend and for the indicators, unstable relationships and sometimes wrong slope signs. Beyond these contrasting results that could be partly due to the quality of yield statistics or the relevance of the selected indicators, CST combined with the SPIRITS tool for extracting indicators at region level from raster time series, should help crop analysts predict crop yield, in particular where many indicators derived from remote sensing data or crop models are to be tested.