Optimized Ensemble Deep Learning for Automated Crop Area Mapping Using Satellite Imagery

Accurate and efficient mapping of crop areas is crucial for optimizing resource allocation and supporting data-driven agricultural decision-making. For crop area mapping, a number of sophisticated Machine Learning(ML) and Deep Learning(DL) techniques have been proposed; however, many of these techniques failed to detect small or irregularly shaped fields just because of the lack of sufficient data and lack of methods for automatic feature extraction. To address these challenges, this study presents a Five-Phase Optimized Ensemble Deep Learning (5P-OEDL) model. The framework is divided into five phases. In the first phase, satellite images are gathered from Google Earth Engine (GEE), focusing on the Kasaragod region of Kerala, India. The images are augmented through various techniques and prepared for subsequent model performance evaluation in the second phase. The third phase involves selecting and fine-tuning three pre-trained Convolutional Neural Networks (CNNs): RPNet34, GLNet152, and ResNet152 known for their feature extraction capabilities in remote sensing applications. These models are fine-tuned on the preprocessed dataset to enhance their ability to recognize crop types in complex agricultural landscapes. In the fourth phase, proportional weights are assigned based on the model accuracy. Finally, the outputs of the three models are combined in the fifth phase using a weighted averaging technique, producing highly accurate crop area maps. This approach is proven to outperform individual models, even with low-resolution satellite images, offering a scalable solution for real-time crop monitoring and resource allocation in Kasaragod and other regions with similar agricultural challenges.