Detection of Different Crop Growth Stages by Applying Deep Learning Over Sentinel-2 Images of Bahía, Brazil

Abstract

This study delves into the agricultural landscape of Bahia, Brazil, employing the Mask R-CNN deep learning model with satellite imagery to detect three crop growth stages (early, mid-growth and maturity stage). This model is suited to the region’s complex terrain and diverse crop patterns, providing accurate instance segmentation crucial for monitoring crop development. Remarkable results have been achieved with a limited dataset of just 54 images for training, underscoring the model’s efficiency in scenarios where extensive data collection is challenging. The validation metric chosen for this study is the Intersection over Union (IoU), preferred for its ability to quantify the pixel-wise overlap between the predicted and actual segmentations, offering a clear measure of accuracy in spatial contexts. An IoU of 90% was obtained, demonstrating Mask R-CNN’s robustness and potential for precision agriculture in challenging environments.