Mapping the intra-annual cropping frequency of African irrigation schemes by multi-sensor fusion

Abstract:

Governments and engineers have promoted the construction of large-scale, formalised, irrigation schemes across Africa for nearly 100 years. These developments are designed to increase food production and reduce vulnerability to climate shocks. Yet over the past decades, many irrigation schemes have deteriorated or completely failed; due to a wide range of problems from faulty infrastructure to unexpected climate shocks. Understanding the drivers of successes and challenges on irrigation schemes is complicated by limited long-term data, while the Landsat archive remains poor for large areas of Africa preventing remote sensing-based alternatives.

Here, we demonstrate a multi-sensor fusion methodology to derive intra-annual cropped area estimates for African irrigation schemes in the 21^st century. Our methodology produces monthly Landsat-like images from the fusion of Landsat 5, Landsat 7 SLC-off, and MODIS imagery, which are classified into cropped area estimates. First, we use the StarFM fusion algorithm to generate monthly Landsat-like images from MODIS composites, based on temporally co-located MODIS and cloud free Landsat 5 or Landsat 7 SLC-on images. Next, we adjust these Landsat-like images against Landsat 7 SLC-off pixels by iteratively reweighting within a spatiotemporal Generalised Additive Model. Finally, we classify the derived monthly, Landsat-like, time-series data using a Random Forest classification model, mapping the number of crop harvested per year for the 2000-2019 period.

We test this methodology against two irrigation schemes in West Africa: the Office du Niger scheme in Mali and the Tono Irrigation Scheme in northern Ghana. For both sites, the mapped areas correlate with official statistics on cropped areas. Our data highlight infrastructure improvement and expansion on the Office du Niger, and the resilience of the scheme to rainfall variability. Whilst on the Tono scheme, we show a vulnerability to large rainfall deficits, and a decline in cropping frequency at scheme edges due to infrastructure deterioration. This methodology is applicable to many areas where the Landsat archive is limited, but intra-annual mapping is required.