Detecting crop residues burning using Sentinel-2 imagery : conservation agriculture promotion in Central Malawi

Malawi is a country which struggles with food security and where a performant agriculture is particularly important. NGO’s and the government have a history of developing projects and initiatives to try to improve agriculture’s productivity. The most widely promoted technique in Malawi historically has been conservation agriculture (CA). The question we scrutinized through this study was that of one particular practice under this umbrella, namely crop residues retention. This practice conflicts with the residues burning tradition, which we chose to observe via remote sensing. The study area was the Central Region of Malawi and specifically the three conservation agriculture promotion sites of Zidyana, Mwansambo and Salima. We focused on the residues burning season of 2016: from early August to November. We have tested three pixel-wise classification methods based on Sentinel-2 satellite imagery. The first one was an expert-based classification combining three different burn indices. The second was a maximum likelihood classification and the last one was a random forest algorithm set out through the Sentinel-2 for Agriculture system. We then calculated statistics of the performance of those methods. The maximum likelihood method showed the highest accuracies, all over 97% (except for an outlier dates). Random forest showed a good performance as well, whereas the expert-based classification was rather inconsistent with values ranging from 68 to 99%. The distribution of burned fields across space and time was explored, which allowed calculating numbers and densities of burned fields. Seasonally, burning occurs more rarely in August and reaches a peak mostly in September and October and is less practiced in November. Spatially, our analysis has shown a higher prevalence of the burning practice in the Salima region. For all sites, the trend is for the number to increase with distance to the promotion site. We could infer that promotion of CA has been effective in Central Malawi, more so in the more established promotion sites of Zidyana and Mwansambo. Time dynamics have shown that burning occurs mainly at the end of the dry season. It is less common the more we approach the humid season’s beginning, where harvest does not take place anymore. Nonetheless, our method to obtain field numbers and densities is flawed and possibly overestimate the number which makes it hard to draw any conclusions. We can in any case rely on the visual interpretation of burned areas maps obtained with the three methods. This master thesis is a proof that remote sensing can be a valuable tool in assessing CA practices adoption. UAV is without a doubt even more accurate and useful in this context than satellite imagery. However, contrary to our hypothesis, residues burning is not necessarily a suitable proxy to grasp adoption of crop residue retention in the landscape.