New ways to use remote sensing phenology and machine learning for predicting irrigated and rainfed agriculture in Africa
Tobias Landmann, David Eidmann, Natalie Cornish, Jonas Franke, Stefan Siebert
Remote Sensing Solutions GmbH, Germany. Technical University of Darmstadt, Germany. Department of Crop Sciences, University of Goettingen, Germany
In this contribution, we present a satellite-based farming system (irrigated or rainfed) monitoring approach
for Zimbabwe that uses optimized harmonics (phenology) from 30-meter Landsat vegetation index
observations (2013 to 2018). Information about the extent of agricultural land and the distinction into
rainfed and irrigated land are needed for many applications such as assessments of agricultural productivity,
food supply, biodiversity or effects of extreme events such as drought. In using harmonic curve parameters,
we aimed to maximize the use of time-series information for accurate and effective farming systems
mapping in tropical regions. Given stable inter-annual seasonality of land use, we found that optimized
harmonics provide an effective way to compute vegetation phenology while accounting for data gaps and
residual noise inherent in Landsat time-series data. Farming systems in Zimbabwe could be mapped with
an overall accuracy of 97% using random forest as a machine learning classification tool.
Event: Annual World Bank Conference on Land and Poverty 2019
Only personal, non-commercial use of this document is allowed.