Hybrid CNN-LSTM Model for Crop Yield Prediction using Drone Imagery

Unmanned aerial vehicle (UAV) based remote sensing has become a valuable tool in precision agriculture, offering high-resolution, timely imagery that supports detailed crop monitoring and yield estimation. Traditional methods relying on satellite data often face limitations in spatial and temporal resolution, which UAVs are able to overcome, providing field-level insights crucial for better agricultural management. Alongside these technological advancements, deep learning methods have emerged as powerful approaches for modeling complex spatial patterns present in agricultural landscapes. In this study, we utilized UAV-captured RGB imagery collected from five distinct agricultural fields near Kurukshetra. Each crop is represented by three time-series images corresponding to the start, middle and end of its growth cycle. For yield prediction, we first trained convolutional neural networks (CNNs) including ResNet50, DenseNet121, and EfficientNet on the UAV imagery. To incorporate temporal dynamics, we further developed hybrid CNN-LSTM architectures, combining DenseNet121, ResNet50 with LSTM layers to model sequential dependencies. Model performance is evaluated using mean absolute error (MAE) and R2 (R-sqaure) as the primary metric. Among all models, the LSTM-DenseNet121 architecture achieved the best performance, obtaining a MAE 20.903 and R2 score of 0.7339 on the test dataset, demonstrating the effectiveness of combining CNNs with LSTMs for accurate crop yield prediction using UAV imagery.