Deep learning system for detection and classification of banana and plantain cultivation zones in satellite images, implementing data parallelism

This work presents the development and implementation of two deep learning approaches for identifying and classifying plantain and banana crop areas from high-resolution satellite imagery. Two main methodologies were compared: A pixel-based classification model using semantic segmentation PSPNet architecture with ResNet-34 and an object-based classification model using the Mask R-CNN architecture with ResNet-50. The pixel-based model showed superior performance in spatial precision and species differentiation, achieving strong metrics ($\text{Dice}=0.762$, $\text{Accuracy}=0.936$) and producing segmentation more consistent with the actual geometry of crop parcels. In contrast, the object-based model reached a mAP of 0.6292 in its best configuration, offering structured detections but with lower accuracy in irregular boundary areas. Both models were trained on high-resolution orthoimages, and their results were evaluated both qualitatively and quantitatively. Additionally, the impact of network architecture, generalization capacity, and computational efficiency was analyzed, considering the role of hardware in training performance. The developed system demonstrates the feasibility of computer vision in precision agriculture and provides a strong foundation for future research in tropical crop monitoring.