The increased temporal coverage afforded by the ESA Sentinel systems provides new opportunities for high temporal frequency moderate resolution remote sensing, enabling a new generation of data products to be generated for the study of land cover and land use. Although agricultural applications provided a strong rationale for the Landsat missions, the 16-day coverage is a serious limitation for obtaining cloud-free observations. Landsat-8 in combination with Sentinel-2 increases the opportunity for cloud-free optical data, and Sentinel-1 provides freely available moderate resolution microwave coverage and an unprecedented opportunity for new products for agricultural land use.
Building on experience gained with MODIS, this proposal prototypes a new crop yield product for wheat, corn and soybean, which can be generally applied to major crop growing regions. We will utilize robust features from multi-temporal optical and microwave imagery and generalized classification models for crop mapping to generate crop specific masks. A-priori knowledge on crop calendars and meteorological data influencing crop growth (e.g. growing degree days) will be used to run and test generalized classifiers that will be applicable to multiple agriculture conditions, providing the possibility of generating dedicated products at continental to global scales. Crop yield mapping and assessment will be performed by first building generalized crop yield models using coarse resolution data (MODIS) that have a long data record (from 2001), and then extrapolating the yield models to moderate resolution data (Landsat-8 and Sentinel-2) to provide yield maps at 30m. Multiple satellite-derived features, such as vegetation indices and biophysical parameters estimated at a single date or accumulated over the crop growth period, will be analyzed, with an NDVI-based approach serving as a benchmark, to connect them with crop yield at regional and field scales to target a 5-10% error. Since temperature is a primary factor affecting the rate of crop development, meteorological data and photosynthetically active radiation (PAR) will be incorporated into the yield models. We will advance the current science by estimating uncertainties of crop yield estimates, as previous approaches usually do not report associated uncertainties. For this, we will utilize non-linear regression models based on Gaussian Processes (GPs) that produce both the estimate and its uncertainty. Incorporation of microwave data from Sentinel-1 will be performed by inversion of the radar signal in multiple polarizations to the biophysical parameters, using the semi-empirical Water Cloud Model (WCM). SARderived parameters, such as LAI, can be used as a basis for integrating with optical-derived parameters used for crop yield mapping.
Crop type and crop yield maps will be generated for administrative regions (with area ranging from 28,000 to 308,000 sq. km) for 7 countries (Argentina, Canada, France, South Africa, Tanzania, Ukraine and US). We will validate these regional products at regional and field scales for test sites in these countries, exhibiting different agriculture practices and conditions (field sizes: 0.5 ha to 500 ha, yield range: 1.5 t/ha to 10 t/ha) where the three major crops are grown and where we have validation data and strong partner collaboration. The project will work with international collaborators who are actively involved in crop monitoring, using Landsat-8 and Sentinel-1/2 data. The international collaboration will focus on the evaluation and validation of the products in the framework of the GEOGLAM program. The new products of crop maps and crop yield maps will have a number of potential users from local authorities dealing with food security, farmers to address agricultural management practices (e.g. yield gaps) and insurance companies concerned with farm yield insurance contracts.
Principal Investigator
Project Sponsor