Effect of using crop specific masks on earth observation based crop yield forecasting across Canada

Abstract

Land cover maps are often required in Earth Observation (EO) data analysis to isolate regions where specific land classes are present. They are normally derived from remote sensing images and ground truthed inputs. The crop cover maps that target specific crop classes – herein referred to as “crop masks” – could be used to identify the pixels that represent the targeted class, which in-turn could potentially improve the inputs for agricultural applications such as crop-specific yield forecasting. In this study, a set of crop-specific masks derived from Agriculture and Agri-Food Canada's (AAFC) annual space-based crop inventory was evaluated for use in forecasting the regional and national yields of five major crops (spring wheat, barley, canola, corn for grain and soybean). Yield forecast results from these masks were compared with the generalized masking approach used operationally by AAFC and Statistics Canada (STC). This comparison was undertaken using two model performance indicators, the Bravais and Pearson coefficient of determination and the root mean squared error, computed using a Leave-One-Out-Cross-Validation (LOOCV) procedure. Yield modelling was conducted at the Census Agricultural Region (CAR), and then aggregated to the provincial and national scales. Overall, an improvement in yield forecasting skill was observed when crop-specific masks were used for all crops in certain regions and at certain spatial scales. However, only the corn and soybean mask showed consistent improvement across spatial scales for most crop growing regions. For spring wheat, canola and barley, improvements were mostly observed in regions where crops are sparsely distributed or clustered. The results of this study will provide further guidance in developing and refining the current crop-specific masks for use in yield forecasting estimates.