Detecting crop damage using Sentinel-2 imagery in a smallholder agriculture landscape

Smallholder farmers in Sub-Sharan African countries are facing more and more challenges in terms of agriculture. Not only the lack of mechanization remains a problem to feed the population that continues to grow but in addition, external threats such as the invasive pest Fall armyworm and climate hazards increase the risk of food insecurity and famine in some countries. Determining the extent of such damage requires extensive field work and a lot of time when it is necessary to act quickly to deliver food aid. Remote sensing can represent a real solution to this problem. Indeed, monitoring crops remotely is a technique that has been used successfully for several years now. The recent availability of free of access Sentinel-2 imagery opens the door to these kind of technologies for any developing countries. Therefore, this master thesis aims, in collaboration with CIMMYT who is actively working on these matters in Zimbabwe, to assess the feasibility of detecting damage on maize in two districts of the country using satellite imagery. An intermediate objective is to produce a crop types map over the area of study. Firstly, we looked at defoliation damage due Fall Armyworm. It turns out that those damage are rather patchy within a field. This in combination with a complex landscape, one must conclude that Sentinel-2 does not have the spatial characteristic to overcome the configuration of these smallholder fields that, on top of the already existing confounding factors, suffered from intense drought. Secondly, this work took a turn after Cyclone Idai hit our study area. Chipinge being badly impacted by extensive floods and heavy winds, it was decided to evaluate Sentinel-2 capacity in detecting the resulting damage. As results, it appears possible to detect flood damage. Indeed, some features, namely the GCI, NDWI2, the NIR band and the CIRE that uses a band in the red-edge appears to be able to differentiate the extreme classes of damage. Furthermore, by comparing Sentinel-2 images before and after the Cyclone with the SWIR band, the areas that got flooded could clearly be distinguished. However, there were some inadequacies between what we could visually see and the ground truth data which highlights the great challenges when it comes to using remote sensing in a complex environment. This work therefore highlights the difficulties that must be overcome in order to be able to detect crop damage remotely in a context of small-scale agriculture. Using a pixel-based Random Forest classification, a cropland mask was obtained with an F-score for crop of 61% and an overall accuracy of 79%. Once the crop pixels were isolated, the classification of maize, sorghum and other crops was performed. Out of the performance rate of 61%, the three classes were classified with an F-score of 72%, 65% and 99% respectively. Finally, a crop types map was produced using VHR Pléiades imagery and the F-score for maize and sorghum was of 63% for Ward 18 and 48% for Ward 16. In both cases, about 50% of crop misclassification occurred with shrublands and grasslands. Both sensors show limitations and these issues might be highly related with the characteristics of the fields in Zimbabwe that are of small size and rainfed, plus the unusual climatic situation on this crop season. This work could therefore underline the challenges faced by remote sensing in smallholder agriculture in part of Zimbabwe and highlight perspectives for further investigations.