Crop yield estimation using agronomic neural network

Abstract

Systems and method for computing yield values through a neural network from a plurality of different data inputs are disclosed. In an embodiment, a server computer system receives a particular dataset relating to one or more agricultural fields wherein the particular data set comprises particular crop identification data, particular environmental data, and particular management practice data. Using a first neural network, the server computer system computes a crop identification effect on crop yield from the particular crop identification data. Using a second neural network, the server computer system computes an environmental effect on crop yield from the particular environmental data. Using a third neural network, the server computer system computes a management practice effect on crop yield from the management practice data. Using a master neural network, the server computer system computes one or more predicted yield values from the crop identification effect on crop yield, the environmental effect on crop yield, and the management practice effect on crop yield.

What is claimed is: 
     
         1 . A system comprising:
 one or more processors;   a memory coupled to the one or more processors and storing sequences of instructions which, when executed by the one or more processors, causes performing:   storing a master neural network, the master neural network trained to compute a past yield effect embedding from one or more past yield maps and to compute predicted yield values from at least the past yield effect embedding and one or more other values;   receiving a particular dataset relating to one or more agricultural fields, wherein the particular dataset comprises one or more past yield maps for a particular agricultural field;   using the master neural network, computing one or more predicted yield values for the agricultural field by at least computing the past yield effect embedding from the one or more past yield maps and computing the one or more predicted yield values from at least the past yield effect embedding and the one or more other values.   
     
     
         2 . The system of  claim 1 :
 wherein the master neural network is further configured to compute a crop identification embedding from crop identification data;   wherein the particular dataset further comprises particular crop identification data; and   wherein the one or more other values comprise the crop identification embedding computed from the particular crop identification data.   
     
     
         3 . The system of  claim 2 :
 wherein the particular crop identification data comprises one or more genome sequences for one or more crops corresponding to the particular dataset;   wherein the master neural network computes the crop identification embedding using a recurrent neural network configured to identify portions of the genome sequences that are correlated to effects on crop yield data.   
     
     
         4 . The system of  claim 3 , wherein the recurrent neural network is a long short-term memory neural network. 
     
     
         5 . The system of  claim 3 , wherein the recurrent neural network is a gated recurrent units neural network. 
     
     
         6 . The system of  claim 1 , wherein the one or more predicted yield values comprises one or more of a risk adjusted yield value, a total profits value, or a crop quality value. 
     
     
         7 . The system of  claim 1 :
 wherein the master neural network is further configured to compute an environmental embedding from environmental data;   wherein the particular dataset further comprises particular environmental data; and   wherein the one or more other values comprise the environmental embedding computed from the particular environmental data.   
     
     
         8 . The system of  claim 7 :
 wherein the particular environmental data comprises one or more time series of predicted weather events and one or more spatial maps of soil properties;   wherein the computes the environmental embedding using a recurrent neural network for weather events and a convolution neural network for soil properties.   
     
     
         9 . The system of  claim 1 :
 wherein the master neural network is further configured to compute an in-season image embedding from in-season images of the agricultural field;   wherein the particular dataset further comprises one or more in-season images of the one or more agricultural fields at particular periods of a growing season;   wherein the one or more values comprise the in-season image imbedding computed from the one or more in-season images of the one or more agricultural fields.   
     
     
         10 . The system of  claim 1 :
 wherein the master neural network is further configured to compute a management practice embedding from management practice data;   wherein the particular dataset further comprises particular management practice data; and   wherein the one or more other values comprise the management practice embedding computed from the management practice data.   
     
     
         11 . A method comprising:
 storing a master neural network, the master neural network trained to compute a past yield effect embedding from one or more past yield maps and to compute predicted yield values from at least the past yield effect embedding and one or more other values;   receiving, at a server computing system, a particular dataset relating to one or more agricultural fields, wherein the particular dataset comprises one or more past yield maps for a particular agricultural field;   using the master neural network, computing one or more predicted yield values for the agricultural field by at least computing the past yield effect embedding from the one or more past yield maps and computing the one or more predicted yield values from at least the past yield effect embedding and the one or more other values.   
     
     
         12 . The method of  claim 11 :
 wherein the master neural network is further configured to compute a crop identification embedding from crop identification data;   wherein the particular dataset further comprises particular crop identification data; and   wherein the one or more other values comprise the crop identification embedding computed from the particular crop identification data.   
     
     
         13 . The method of  claim 12 :
 wherein the particular crop identification data comprises one or more genome sequences for one or more crops corresponding to the particular dataset;   wherein the master neural network computes the crop identification embedding using a recurrent neural network configured to identify portions of the genome sequences that are correlated to effects on crop yield data.   
     
     
         14 . The method of  claim 13 , wherein the recurrent neural network is a long short-term memory neural network. 
     
     
         15 . The method of  claim 13 , wherein the recurrent neural network is a gated recurrent units neural network. 
     
     
         16 . The method of  claim 11 , wherein the one or more predicted yield values comprises one or more of a risk adjusted yield value, a total profits value, or a crop quality value. 
     
     
         17 . The method of  claim 11 :
 wherein the master neural network is further configured to compute an environmental embedding from environmental data;   wherein the particular dataset further comprises particular environmental data; and   wherein the one or more other values comprise the environmental embedding computed from the particular environmental data.   
     
     
         18 . The method of  claim 17 :
 wherein the particular environmental data comprises one or more time series of predicted weather events and one or more spatial maps of soil properties;   wherein the computes the environmental embedding using a recurrent neural network for weather events and a convolution neural network for soil properties.   
     
     
         19 . The method of  claim 11 :
 wherein the master neural network is further configured to compute an in-season image embedding from in-season images of the agricultural field;   wherein the particular dataset further comprises one or more in-season images of the one or more agricultural fields at particular periods of a growing season;   wherein the one or more values comprise the in-season image imbedding computed from the one or more in-season images of the one or more agricultural fields.   
     
     
         20 . The method of  claim 11 :
 wherein the master neural network is further configured to compute a management practice embedding from management practice data;   wherein the particular dataset further comprises particular management practice data; and   wherein the one or more other values comprise the management practice embedding computed from the management practice data.