Semantic Segmentation of Land Covers Using Deep Learning: A Case Study in the Franciacorta Wine-growing Area

Land cover mapping is a fundamental tool for understanding land use patterns globally. It helps in monitoring changes, assessing ecosystem health, and supporting conservation efforts. The advent of satellite imagery and advancements in remote sensing technology have significantly improved land cover mapping. However, challenges such as managing large datasets, high costs of specialized data acquisition, and the need for optimal deep learning models remain. This article explores a novel approach to semantic segmentation of land covers using deep learning models with pre-trained backbones, focusing on the Franciacorta wine-growing area in Italy.

Researchers Girma Tariku*, Isabella Ghiglieno, Andres Sanchez Morchio, Luca Facciano, Celine Birolleau, Anna Simonetto, Ivan Serina, and Gianni Gilioli from the University of Brescia conducted this study. They introduced an efficient methodology for semantic segmentation in agricultural areas using satellite images and deep learning models with pre-trained backbones. The study addresses gaps in the existing literature by presenting a comprehensive method for dataset preparation and model evaluation.

The study employs three cutting-edge deep learning-based segmentation models: U-Net, SegNet, and DeepLabV3. Each model was tested with various pre-trained backbones, including ResNet, Inception, DenseNet, and EfficientNet. The dataset, named “Land Cover Aerial Imagery” (LICAID), was manually annotated for seven distinct land cover classes: grasslands, arable land, herb-dominated habitats, hedgerows, vineyards, tree-dominated man-made habitats, and Olea europaea groves.

Satellite imagery was acquired using Google Earth Pro and georeferenced with ArcGIS software. The images underwent multiresolution segmentation with eCognition software, and a plant expert validated the segmented shape files using QGIS. The images and masks were then resized and patchified for training the segmentation models.

Models and Backbones

1. U-Net: A deep learning model with a symmetrical encoder-decoder structure, designed for pixel-wise classification tasks.
2. SegNet: Features an encoder-decoder architecture optimized for semantic segmentation tasks.
3. DeepLabV3: Incorporates atrous convolution for multi-scale contextual information and efficient up-sampling for high-resolution segmentation maps.

Each model was trained with and without a backbone, comparing performance metrics such as accuracy, precision, recall, F1 score, Jaccard Coefficient (IoU), and Mean IoU.

The study found that incorporating pre-trained backbones significantly improved model performance. Among the tested models, DeepLabV3 with an EfficientNetB0 backbone achieved the highest accuracy (76.33%), precision (76.10%), recall (76.30%), and F1 score (75.60%). This model outperformed others in accurately delineating land cover classes.

The research highlights the transformative potential of integrating deep learning techniques with satellite imagery for scalable and efficient environmental monitoring and conservation efforts. By addressing challenges such as data availability and computational resources, this study contributes to advancing the field of remote sensing and supporting sustainable environmental stewardship.

This study presents a cost-effective and efficient methodology for land cover mapping using deep learning models with pre-trained backbones. The findings underscore the importance of leveraging advanced methodologies for accurate and reliable land cover mapping, offering scalable solutions for environmental monitoring and conservation.

For a more in-depth understanding of this study, you can access the full article here. Tariku, G.; Ghiglieno, I.; Morchio, A. S.; Facciano, L.; Birolleau, C.; Simonetto, A.; Serina, I.; Gilioli, G. Semantic Segmentation of Land Covers Using Deep Learning with a Pre-trained Backbone: A Case Study in in the Franciacorta Wine-growing Area. Preprints 2024, 2024061947.

#DeepLearning #LandCoverMapping #SemanticSegmentation #SatelliteImagery #RemoteSensing #AgricultureTech #AIResearch #EnvironmentalMonitoring #ConservationTech #Franciacorta #SustainableLandManagement #GIS #MachineLearning #ImageProcessing