- Marginal lands are now being used for crop production to feed an ever-growing population.
- Agriculture productivity must increase to meet demand, intensification rather than intensification.
- Currently, an urgent need exists for information to maximize yield based on land capabilities to mitigate land degradation, improve productivity, and to alleviate food insecurity.
- NASA's role in Earth observation is essential for merging cutting-edge modeling with very high-resolution commercial data to inform the sustainability of current land use and food security.
- Long-term effects of a lack of land tenure rights can be addressed with multi-resolution remote sensing data to relieve land pressure and move toward sustainable management practices.
How is land cover/use classified?
A base UNet convolutional neural network (CNN) deep learning model developed in Senegal using a spatiotemporal diversified dataset was used throughout the Amhara region of Ethiopia (>200,000 sq. km). Using transfer learning techniques with a few high-quality additional training datasets developed by the LIFT-Ethiopia project and further enhanced by our research project [1,2], an ensemble of 3 different fine-tuned CNN architectures was used on WorldView 2 m resolution imagery [1,3]. Land cover was classified with an overall accuracy > 90%. Shrub/trees, settlements, and croplands were classified with an accuracy > 95% (Figure 1).
Figure 1: Burie city suburbs (2100 m – 2250 m amsl): (bottom-left) location map of the study area; (top-left) 2022-02-04 WV03 imagery (2 m); (top-right) classified map by this research (2 m); (bottom-right) corresponding ESA’s World Cover classified map from Sentinel-2 (10 m).
Why is this Important?
Observational scale matters when mapping sub- hectare fields. Our LCLU product is the best
available in the region and will be beneficial in making informed decisions by the Amhara Region GOs, and NGOs on sustainable land use management, precision agriculture, ecosystem services modeling, climate change mitigation, urban planning, market management, and food security programs. Particularly, the cropland layer output can be used as a baseline for recurrent seasonal yield estimation and forecasts that are useful for planning purposes to alleviate food insecurity.
The way forward?
The CNN models were successfully tested on open-source datasets, such as Planet, Sentinel 2, and Harmonized Landsat Sentinel (HLS). Recurrent model application for seasonal/annual monitoring could be performed by government officials and NGOs in the study area and elsewhere. Technology transfer to concerned sectors, other relevant NGOs, and the wider research community will empower them with information to make sustainable land use
[1] Li et al., (2022). J. Sens., 22(12), 4626. [2] Neigh et al., (2019). IEEE IGARSS 2019, 5397–5400. [3] Défossez wt al., (2020). arXiv:2003.02395.
Project Investigator: Alemu et al, NASA GSFC, MD, USA; Email woubet.g.alemu@nasa.gov
The opinions expressed are solely the PI's and do not reflect NASA's or the US Government's views.