Early crop mapping at continental scales derived from reconstructed high spatial resolution images

Abstract:

High interannual variability in crop production can be attributed to many unmanageable and unpredictable weather related factors which complicate food security and planning. Obtaining crop classification maps before harvest would have a significant impact on our ability to monitor crops and correct emergent anomalies; and would help in administrative actions (i.e loss adjuster reports). Remote sensing offers an objective solution for continuous crop monitoring over large spatial extents, and is a necessary step to improve current satellite-based remote sensing crop yield models. Nowadays there are several crop cover datasets like the Cropland Data Layer (CDL), which provides Landsat-scale annual cropland data and is used generally as an available target crop label. These datasets either are provided earlier the following year and do not reflect the current crops on the ground, or are more frequent, but at coarse resolutions. This work develops an early crop classification map based on the diversity of the different crop phenologies over continental scales at high spatial resolution. To address these objectives, a classification analysis is done covering the contiguous US (CONUS) in 2019 using smoothed and gap-filled Landsat reflectance data. The monthly temporal series of 2019 was generated from the HIghly Scalable Temporal Adaptive Reflectance Fusion Model (HISTARFM) algorithm, which combines the fine spatial resolution of the Landsat TM sensor and the higher temporal resolution of the Moderate Resolution Imaging Spectroradiometer (MODIS), while avoiding spatial gaps due to clouds and aerosols. Furthermore, the CDL is available over the CONUS and is used as a label in the classification. These data are processed with Google Earth Engine, a cloud-based application that can be used to map crops globally and provides several optimized machine learning algorithms like support vector machines or random forests to obtain classification images at broad scales. Finally, it is shown that accurate high spatial resolution crop classification maps can be achieved months earlier than the available options. These early classification maps are highly valuable for society and agriculture and are essential to develop advanced decision support systems to inform better land and water allocation strategies.