Abstract:
Yield prediction is helpful for timely harvest management, crop planning, and food security. It depends on many factors like location, climate, soil characteristics, genotype, etc. The data used in yield prediction is a typical mix of highly dynamic time series (meteorological) and static (soil) data. We effectively integrate the two data categories to train a deep-learning model. We introduce a novel attribute selection algorithm to select the most discriminating soil features and modified it for depth-level selection which suggests the most appropriate depth of soil factors for a given crop. We have also introduced a novel approach for modeling the problem where spatiality is handled by clustering locations based on their meteorological and soil characteristics which allow our model to learn spatial patterns. The variation in sowing and harvesting time across locations is taken care of by using padded crop cycle data. We have also taken several other design decisions and validated them on existing models. We experimented with NC94 data of the US with three major crops soybean, wheat, and corn, and predicted yield at the county-level. We have also modified our model to perform in-season and multi-time horizon prediction. The results of our proposed YieldPredictNet show that it outperforms competing techniques.