Towards a Framework for Monitoring Crop Productivity in Agroforestry Parklands of the Sudano-Sahel Using Sentinel-1 and 2 Time Series

Abstract

The agroforestry parklands in the Sudano-Sahelian zone are of critical importance for food security, but presently face several challenges in terms of changes in climate, land use practices and intensification. The ability to systematically monitor crop productivity in these systems is therefore of importance for both informing land management policies and studying long-term trends. This study, conducted in two different agroecological areas in southern and central Burkina Faso covering two climate-wise very contrasting years (2020-2021), is an initial step to designing a system based on satellite remote sensing that enables national-scale monitoring of crop productivity. While there are several steps before this can be realized in practice, our results provide key insights into topics such as how the field data collection and modeling should be done. The main assessments focused on how to best process and combine remote sensing data sources, including time series from the Sentinel-1 and Sentinel-2 satellite systems, as well as soil properties, elevation and tree cover. Other key focuses were evaluating different regression modelling algorithms (multilinear and machine learning) and clarifying the potential benefits of performing the modeling in specific geographic regions and years or if the modeling can be generalized. Overall, the results show that accurate estimates of crop productivity are achievable using the proposed modeling framework, with encouragingly high R2 (0.65-0.80) and low root mean square errors (12-20%). Sentinel-2 was the most important data source, but our results also demonstrate the potential of Sentinel-1, which has the benefit of not being affected by clouds. Another encouraging aspect is that the results were stable both between the years, which differed significantly in terms of rainfall and crop productivity, and between the sites that are characterized by contrasting crop compositions. This study shows that the development of a national-level crop monitoring system in Burkina Faso or countries with similar environmental conditions is within reach. Open access remote sensing data sources and cloud-based data processing are potential alternatives for improving generalization and consistent monitoring over the years. A precondition for such a system to work, however, is the availability of high-quality reference data that captures variability in climate and crop conditions. One way to achieve this could be through the development of crowdsourcing initiatives, where local farmers could be trained to collect and provide crop surveys across regions.

Keywords: Crop productivity, agroforestry, Satellite remote sensing, Machine learning, Sentinel 1-2, parkland