In-Season Growth Forecasting in Cotton Using Unmanned Aerial System- based Canopy Attributes and LSTM Models

Abstract

Cotton (Gossypium spp.) is one of the important cash crops in the United States. Monitoring in-season growth metrics, from early season growth to harvest, is crucial for predictive and prescriptive cotton farming. In recent years, forecasting models have garnered considerable attention to predict canopy indices. This allows selection of management options during crop growth to boost cotton yield and profitability. Here, we used unmanned aerial system-derived canopy features, including canopy cover, canopy height, and excess green index, collected from 3500 plots at Driscoll in Corpus Christi, Texas during the years 2019, 2020, and 2021 for in-season growth forecasting. Training datasets in our model were produced by K-Means clustering and Dynamic Time Warping (DTW) techniques were used to compare various Long Short-Term Memory (LSTM) models in predicting the three canopy features. Accuracy was determined using Root Mean Square Error (RMSE). Results indicated higher predictive capacity of Convolutional Neural Networks (CNN) LSTM for canopy cover, and multi-layer stacked LSTMs for canopy height and excess green index respectively. Overall, results show tremendous potential for in-season growth forecasting and management of agricultural inputs like pesticides and fertilizers for improving crop health and productivity.