Assessing cropland disagreement in Tanzania using machine learning methods with Sentinel-2 and Planet Scope imagery

ABSTRACT

East Africa faces major land use pressures arising from the harsh effects of climate change and fast-increasing human populations. Recent efforts to increase food production in this region have focused on cropland mapping and approaches to improve yield. Generating accurate cropland maps provides critical support to inform policy, investment, and logistical decisions that address food security, but cropland disagreement assessment studies are lacking mostly due to the unavailability of fine-scale data. Understanding the inconsistent nature of pre-existing global cropland mapping products necessitates disagreement assessment as the most valuable urgent tool to align agricultural resource allocation and policy realignment to tackle food security challenges. This study evaluates cropland disagreement in Kilombero and Ulanga districts in Morogoro, Tanzania based on land use land cover (LULC) maps generated using machine learning methods and high-resolution PlanetScope and Sentinel-2 data sets acquired from 1 January 2017 to 31 December 2018. We created time-variant median composites and performed a three-step assessment; (i) generating multi-class LULC maps in Google Earth Engine, (ii) generating croplands and non-cropland LULC classes from the multi-class images, and (iii) assessing cropland disagreement based on GIS conditional statements. Results show minimal disagreement in mapping cropland based on higher-resolution PlanetScope and Sentinel-2. Results based on the Random Forests and Support Vector Machines relatively outperform those from Classification and Regression Trees achieving 93%, 89%, and 83% with PlanetScope and 91%, 86%, and 76% with Sentinel-2, respectively. On average, 73% of the pixels were consistently classified while 27% were misclassified between cropland and non-cropland classes. Prior unsupervised land cover cluster analysis may substantiate the quality of sampling for accurate land cover classification. Our results indicate that in data-sparse regions where confirmatory ground truth data are lacking, remotely sensed cropland identification may be more successful using Support Vector Machines on newer high-resolution imagery, thereby improving crop forecasts.

KEYWORDS:

• PlanetScope
• Google Earth Engine
• machine learning
• binary classification
• Sentinel-2
• East Africa

Acknowledgements

The authors would like to extend their appreciation to their scientific networks for their support and help during the development of this work and field data validation. Special thanks to Dr Chengxiu Li whose previous work has greatly inspired this work.

Disclosure statement

No potential conflict of interest was reported by the authors.