Researchers from Zhejiang University and Tongdun Technology, both located in Hangzhou, China, have utilized deep-learning methods to enhance crop-yield predictions. By considering the impact of farmland location, these techniques offer more accurate predictions for farmers and policymakers. Accurate crop-yield predictions are crucial for agricultural success and effective decision-making. With climate change and the need for increased food production, precision in these predictions is more important than ever.
In the past, predicting crop yields relied on tracking factors such as weather and soil conditions. While many of the variables used in these predictions remain the same, recent advancements in modeling techniques have made them more sophisticated. Deep-learning techniques now have the ability to not only calculate the individual effects of variables like precipitation and temperature on crop yield, but also the interaction between these variables. The interplay between different variables can lead to different results compared to analyzing them independently.
The researchers employed a recurrent neural network, which is a tool that tracks the relationships between variables over time, to capture the complex temporal dependencies that impact crop yield. According to Chao Wu, a researcher at Zhejiang University and one of the authors of the paper, variables such as temperature, sunlight, and precipitation change over time and interact with each other in complex ways, ultimately affecting crop yield cumulatively. This model also accounted for variables that are difficult to quantify, such as advancements in breeding and agricultural cultivation techniques, allowing it to capture larger trends that extend beyond a single year.
To incorporate spatial and regional information into their predictions, the researchers combined their recurrent neural network with a graph neural network representing geographic distance. This addition enabled the model to consider the proximity between different regions of farmland and assess how predictions for specific locations would be influenced by the surrounding areas. By including information about adjacent regions, the researchers were able to leverage relationships across both time and space.
To test the effectiveness of their method, the researchers used U.S soybean yield data provided by the National Agricultural Statistics Service. They inputted various climate, soil, and management data into the model, training it on yield data from 1980 to 2013 and testing it on data from 2015 to 2017. The proposed method outperformed existing models, including non-deep-learning methods and deep-learning models that didn’t account for spatial relationships.
In their future work, the researchers aim to make the training data more dynamic and enhance security features in the model-training process. Currently, the model is trained on aggregated data, which raises concerns about the confidentiality of proprietary information. In the agricultural industry, exposing data like crop yields and farm-management practices could lead to unfair advantages for competitors and make farmers more susceptible to scams and theft. To address these concerns, the researchers propose using a federated learning approach, which would update a global model while keeping different data sources isolated from one another.
By applying deep-learning techniques and accounting for spatial relationships, the research team demonstrated significant improvement in crop-yield predictions.
The ainewsarticles.com article you just read is a brief synopsis; the original article can be found here: Read the Full Article…
