Operational framework to predict field level crop biomass using remote sensing and data driven models

Remote Sensing (RS) based monitoring provides opportunities to acquire timely and reliable information on crop growth at diverse scales. Crop yield forecasts can help decision makers to formulate policies on maintaining national food reserves, sustaining food supply chains and attaining national food security. Inter field crop yield heterogeneity arising from varied field management and agricultural practices necessitates this forecasting to be done at field-level. Such field level information can aid farmers to identify gaps in water and farm management practices and take corrective actions, if needed. So far, no scalable tools are available to predict crop yields at field level that leverage the availability of open-access high-resolution RS data. This research implements an operational framework to predict field level crop biomass by evaluating different regression algorithms to develop data driven models, leveraging historical and near real time open-access high resolution optical satellite data from Sentinel-2, radar data from Sentinel-1 and evapotranspiration (ETa) and Net Primary Production (NPP) data from Food and Agriculture Organization's (FAO) Water Productivity through Open-access Remotely sensed data (WaPOR) platform. NPP is used as a proxy for biomass production/yield. Five of the most commonly used regression algorithms were tested to build a data-driven model for sugarcane NPP prediction in Wonji-Shoa sugarcane estate, located in the Awash Basin, Ethiopia. The models tested were the Multivariate Linear Regression (MLR), Stepwise Multivariate Linear Regression (SMLR), Boosted Regression Trees (BRT), Support Vector Regression (SVR), and Random Forest Regression (RFR). The results revealed that for seasonal sugarcane NPP predictions, the linear regression models (MLR and SMLR) yielded more accurate predictions than the non-linear machine learning models (BRT, SVR and RFR) tested. The highest accuracy was achieved for MLR models for which estimates with 89% accuracy could be made 4 months prior to the harvest and with accuracies of 79% up to 200 days (approx. 6.5 months) before the harvest. The non-linear machine learning models, however, could not provide reliable estimates of sugarcane NPP (accuracies < 61%). Cumulative vegetation indices (VIs) were found to have higher predictive power than standard VIs for predicting future sugarcane NPP. Cumulative Enhanced Vegetation Index (EVI) was found to be the variable with the highest predictive power, followed by VH polarized Sentinel-1 Synthetic Aperture Radar data and WaPOR ETa. The study shows the usefulness of high-resolution RS information to predict seasonal NPP at field level. The methods presented here can be translated into an automated framework towards an operational system.