Annual Crop Yield Predictions using AI and EO

Abstract

Climate change is intensifying hydrometeorological hazards such as droughts, which directly impact agricultural yield and threaten food security for the increasing global population. Agricultural drought, defined as a reduction in crop productivity due to reduced soil moisture, is especially crucial in understanding drought-related impacts on crop yield and predicting these with enough lead time to enable better preparedness, thereby fostering climate resilient food security. Earth Observation presents a unique opportunity to monitor various hydrometeorological, crop, and vegetation health indicators, which could help to simulate drought impacts on crop yield. This study used a Random Forest Regressor to predict crop yield deficits resulting from agricultural droughts. The model is trained using ERA-5 Land monthly and PKU GIMMS NDVI data as inputs, with crop yield data used as targets. The analysis was performed for wheat and barley crops, with lumped/large-scale (single model for all districts) and distributed/fine-scale (individual model per district) implementations. We also analyzed feature importance using the mean decrease in impurity (Gini) to assess which features are important at which scale. The model performed better for wheat than barley in both cases here. The essential features in lumped analysis for barley were NDVI, temperature, radiation, and soil moisture layer 3 (28 -100 cm). For wheat, NDVI, albedo, evaporation, latitude and radiation were important features. In the distributed analysis, the top three features for barley were NDVI, wind speed, and precipitation, while for of wheat, NDVI was the most important for a majority of the districts, followed by evaporation and humidity. This study demonstrates that different factors are important for the predictability of annual agricultural drought induced changes in crop yield at distinct spatial scales and for different crops. The insight gleaned from this analysis pave the way for developing scale-relevant robust crop-specific yield prediction models, which can enable anticipatory action to mitigate drought impacts, resulting in tangible societal benefits.

Authors 2

1. Pradeep Kumar Garg

   Indian Institute of Technology Roorkee

   Affiliation as printed

   Indian Institute of Technology (IIT Roorkee),Department of Civil Engineering,Roorkee,India

2. Antara Dasgupta Aachen Institute of Hydraulic Engineering and Water Resources Management

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Institute of Hydraulic Engineering and Water Resources Management (IWW),Aachen,Germany