Crop yield prediction method based on multi-modal and dynamic weight optimization stack integration

Abstract

The application discloses a crop yield prediction method based on multi-modal and dynamic weight optimization stack integration, which comprises the following steps: preprocessing and feature fusion are performed on multi-modal data to obtain multi-modal fusion features, the multi-modal fusion features are input into multiple base models, a time sequence input vector sequence is constructed and input into a gated recurrent unit network by using a yield preliminary prediction sequence and an environmental context feature corresponding to each time step in the yield preliminary prediction sequence, a dynamic weight is obtained, at each time step, a dynamic weighted yield prediction sequence is formed by weighted summation according to the dynamic weight sequence, and then the dynamic weighted yield prediction sequence is input into a ridge regression meta-learner for learning, and finally a crop yield final prediction value is output. The application adopts a stack architecture of various base models, dynamically generates integrated weights adaptive to environmental changes by using a gated recurrent unit network, and finally corrects and outputs by a ridge regression meta-learner, so that high-precision and self-adaptive stable prediction of crop yield under variable climate and production conditions is realized.

Images