Cotton is one of the important agricultural products and a classic cash crop supporting both the American and worldwide economy. The United States is the third-largest cotton producer after China and India. Approximately 17.62 million cotton bales were produced in the US in 2021. Texas is a leading producer of cotton with approximately 7 million acres planted of the total 12.2 million cotton acres planted in 2022. The United States cotton yield is projected around 16.5 million bales in 20221, and cotton farming generates more than 125,000 jobs and more than $21 billion economic output in the rural southern areas annually 2 .
Genetic, environmental, and managemental impacts are often reflected on cotton plants’ growth and development. Multifold sensors are commercially available to understand such responses. In recent years, Unmanned Aerial Systems (UAS) with RGB and multispectral sensors are increasingly used in monitoring plant growth. Canopy spectral reflectance and leaf area indices are important cotton attributes used in forecasting plant growth and production under irrigation management. Zhao et al3 highlighted the importance of reflectance data and other agronomic attributes acquired during early flowering for yield prediction. UAS are often advantageous in monitoring reflectance characteristics more rapidly and in real-time. On a commercial farm in Australia, UAS-derived vegetation indices were used to forecast in-season nitrogen status and lint yield4. Models could be developed based on UAS-derived data, which allows canopy parameters forecasting and crop management decisions including irrigation, growth regulators, fertilizers, and harvest-aid applications. Indeed, UAS coupled with artificial intelligence and big data analytics enables better understanding of plant growth5, 6 and forecasting of growth factors for genotype selection. However, there remains a significant technology gap in the ability to forecast crop growth and production during the cropping season; thereby affecting in-season management choices. As such, machine learning (ML)-based plant growth models hold promise for in-season growth forecasting, crop management, and precision farming for improving agricultural production and profitability through informed decision-making7.
Process-based simulation models are commonly used models in crop management, but they require substantial calibration. Therefore, ML and in particular, deep learning neural network models like Long Short-Term Memory (LSTM)8 networks could be among the portfolio of options available to improve the performance of computational models. These models may further benefit from using canopy metrics, such as canopy cover, canopy height, and excess green index, with the highest degree of accuracy beginning in the early part of the growing season for yield forecasting. Few studies are also available on high resolution UAS data to determine biologically relevant plant canopy parameters and use of Artificial Neural Network (ANN) in forecasting cotton yield9.
Recurrent neural networks (RNNs) that can recognize long-term dependencies belong to the class of LSTMs. They do not have difficulty in picking up new material; in fact, it's almost like it comes naturally for them to retain the material for a long time. Multiple studies have compared the effectiveness of different LSTM models for predicting canopy features, including vanilla LSTM, bidirectional LSTM10, 11, stacked LSTM12, 13, and CNN-LSTM14, 15, 16. Such studies have evaluated the usefulness of several variants of LSTM models in predicting growth metrics in cotton crops. Stacked LSTM is peculiar as it uses many stacked LSTM layers, all of which are linked together. Due to the bidirectional LSTM, which trains two input sequences rather than one, the first and second sequences are perfect clones of each other, thus allowing the model to learn more quickly. Bidirectional LSTM refers to the ability for a neural network to remember sequences in either direction, from the future to the past or the other way around (past to future). A bidirectional LSTM is different from a standard LSTM in that it takes input in both directions. Traditional LSTMs only allow for unidirectional input flow, either forward or backward. To build a CNN LSTM, it is necessary to add a CNN layer, then an LSTM layer, and finally a Dense layer. CNN LSTM networks can exploit convolutional structures in both input-to-state and state-to-state transitions for spatiotemporal predictions. The encoder-decoder LSTM is a recurrent neural network used to solve sequence-to-sequence problems (seq2seq problems). Prediction issues arising between sequences can be challenging to solve because input and output sequences may have different item counts. Sequence-to-sequence challenges include text translation and program execution instruction. An LSTM model was employed in this study. Using the root-mean-squared error (RMSE) as a performance metric, the best model for predicting each canopy attribute was explored.
Long Short-Term Memory models have historically been used extensively in forecasting crop yield17, 18, 19; commodity price; and occurrence of pests and diseases20, 21, 22. Due to its non-linear properties and ability to run parallel input series, discover causal linkages, and forecast time series, LSTM models primarily function as predictive models23. In fact, they employ time-series data to boost prediction precision, forecast growth and yield, and support crop management choices. For instance, in one study, five LSTM models—vanilla LSTM, bidirectional LSTM, stacked LSTM, CNN LSTM, and convolutional LSTM—were utilized in forecasting agricultural prices using time series data from five distinct crops. Results indicated a potential for more accurate price forecasting one month in advance. Similarly, the ability to forecast rice pest attacks using bidirectional LSTMs with multivariate time-series input was compared to that of vanilla and stacked LSTM24. Overall, only a few research studies have utilized LSTM to predict the characteristics of crop growth. To the best of our knowledge, no distinct LSTM models or variants have been applied to growth forecasting. To determine the most effective model for forecasting, here, we included and compared many variants of LSTM models to predict canopy cover, canopy height, and excess green indices in cotton crop.