Abstract

Rice is one of the world’s staple food sources, with millions of tonnes produced and consumed every year. Therefore, mapping rice paddies is essential for agricultural management and ensuring food security. This study aimed to classify rice crops in southern Brazil using the SENTINEL-1 SAR time series and deep learning models, comparing two architectures (U-net and LinkNet) and four backbones (ResNet-34, ResNeXt-50, DenseNet-121, and VGG16). The time series construction considered ten images, each for a month, covering the rice planting cycle. The Convolutional Neural Network architectures were adapted to use multi-band data, allowing the extraction of features from all-temporal images. This approach provides capturing spatiotemporal information from rice plantations, which favors its detection. Besides, the research evaluated three data sets considering the polarizations: (a) VV-only, (b) VH-only, and (c) both VV and VH (VV + VH). The classification accuracies used to measure the performance of the models were the overall accuracy, F1-measure, area under the precision-recall curve (AUPRC), and the intersection over union (IoU). Results show that the VH + VV polarization combination yielded the best results, followed by VH-only and VV-only. The VV-only polarization had significantly worst results (nearly 10% less IoU than VH-only and nearly 15% less IoU compared to VV + VH). The results show that rice fields can be successfully classified with deep learning models and through our evaluation the LinkNet architecture with the ResNeXt-50 backbone showed the best results with an accuracy of 0.98, F1 of 0.93, AUPRC of 0.93, and IoU of 0.91.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call