|FENG, AIJING - University Of Missouri
|ZHOU, JIANFENG - University Of Missouri
|Sudduth, Kenneth - Ken
Submitted to: Precision Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/11/2023
Publication Date: 9/5/2023
Citation: Feng, A., Zhou, J., Vories, E.D., Sudduth, K.A. 2023. Prediction of cotton yield based on soil texture, weather conditions and UAV imagery using deep learning. Precision Agriculture. 25:303-326. https://doi.org/10.1007/s11119-023-10069-x.
Interpretive Summary: Cotton yield prediction is important for farmers to conduct proper field management and make marketing decisions. Information on soil texture and weather conditions, supplemented with unmanned aerial vehicle (UAV) imagery, can provide an accurate estimate of yield. Combining images from multiple years can lead to even more accurate predictions. A field study was conducted by ARS scientists and university colleagues from Portageville and Columbia, Missouri, in 2017 through 2019 with the goal to quantify cotton yield variation due to soil texture and weather conditions using UAV imagery and deep learning, an advanced machine learning technique. Soil apparent electrical conductivity provided detailed information about the spatial variation of soil texture within fields, which is relatively stable over time. Unmanned aerial vehicle technology provided information about the spatial and temporal variation of plant properties within fields, and seasonal weather data was also used by the model. Results showed that a model trained with data from two years could accurately predict the cotton yield in a third year. This approach may be useful to researchers and to farmers who are interested in predicting cotton yield at a high spatial resolution.
Technical Abstract: Crop yield prediction is important for farmers to conduct proper field management and make marketing decisions. Yield prediction models that are usually built on a single type of data (e.g., imagery, soil, or weather) may not reflect the holistic effect of environment and management on crop development. The goal of this study was to quantify cotton yield variation due to soil texture and weather conditions using multiple year unmanned aerial vehicle (UAV) imagery and deep learning techniques. UAV images were collected about once a month to quantify cotton growth at two irrigated cotton fields in three years (2017 – 2019). Soil apparent electrical conductivity (ECa) of the fields was measured and calibrated to quantify the spatial variation of soil texture and soil water holding capacity using eleven soil features. A convolutional neural network (CNN) was applied to all the soil features to extract different data transforms of soil data in seven different depths. Similarly, a second CNN was used to analyze the six weather parameters derived from the historical data of a nearby weather station. A gated recurrent unit (GRU) network was used to predict cotton yield using the processed soil data, weather data, and UAV-based image features (e.g., NDVI) of different months. Results show that the GRU model trained with data in two years could predict the cotton yield in a third year with prediction errors of mean average error (MAE) from 247 (8.9%) to 384 kg ha-1 (13.7%). The study indicates that the developed CNN and GRU networks have the potential to predict crop yield of years other than the ones used for training the models.