Identification of Irrigation Canals using Deep Learning Techniques from Remote Sensing Data

The global water crisis has emerged as one of the most pressing challenges of recent decades, particularly in arid regions like Iran, where agriculture accounts for over 85% of water consumption. Efficient water management and conservation are critical to addressing this issue. Irrigation canals are essential for water allocation and reducing the water leakage in agricultural fields require precise and updated mapping to optimize their functionality. While concrete canals typically maintain stable maps, earthen canals experience continuous changes in position and dimensions, necessitating frequent monitoring and updates. Traditional manual mapping methods are time-consuming, costly, and inadequate for large-scale continuous monitoring. To overcome these limitations, this research assembled a dataset of high-resolution satellite images from Khuzestan province, Iran, with detailed labeling of irrigation canals. Semantic image segmentation techniques based on deep learning were then evaluated to identify and map these canals. The result showed a mong the six models tested in this research, the UNet3+ model achieved the highest performance with a recall of 86.64%, F1-score of 89.47%, and IoU of 64.15%. Notably, the model was perfected in detecting narrower and more complex canals. These findings highlight the potential of advanced deep learning models for accurate large-scale mapping of irrigation canals, providing valuable insights for researchers, policymakers and water resource managers in improving irrigation network monitoring and development. Introduction Efficient management of agricultural water resources is a critical challenge in arid and semi-arid regions, where irrigation infrastructure plays a decisive role in sustaining food production and rural livelihoods. Irrigation canals, particularly earthen canals, form extensive networks that distribute water across agricultural landscapes. However, these canals are often poorly mapped, subject to frequent morphological changes due to erosion and sedimentation, and difficult to monitor through conventional field-based surveys. In regions such as Khuzestan Province, where irrigation networks are dense and highly dynamic, the lack of accurate and up-to-date canal maps poses a major constraint for water resource planning, maintenance, and loss reduction. Remote sensing technologies, combined with advances in deep learning, offer powerful tools for large-scale, automated mapping of surface features. Semantic segmentation using convolutional neural networks (CNNs) has demonstrated strong potential for extracting linear and narrow objects such as roads, rivers, and urban infrastructure from high-resolution satellite imagery. Nevertheless, most existing studies have focused on natural water bodies with clear spectral signatures, such as rivers and lakes, while relatively little attention has been given to engineered irrigation canals—especially earthen canals that exhibit spectral similarity to surrounding soils and vegetation. This study addresses this gap by systematically evaluating the performance of six deep learning architectures for the semantic segmentation of agricultural irrigation canals using satellite imagery. The primary objective is to identify the most effective model for accurately detecting canals of varying widths and visual complexity. By benchmarking multiple networks and analyzing their strengths and limitations, this research aims to provide a robust methodological foundation for automated irrigation canal mapping and to support data-driven water management under increasing water scarcity. Methods The methodology of this research is based on a deep learning–driven semantic segmentation framework applied to remote sensing imagery. A dataset of satellite images representing agricultural irrigation canals was collected from Khuzestan Province, covering a wide range of environmental conditions, canal materials (earthen and lined), background textures, and illumination scenarios. The dataset includes both clearly visible canals and visually challenging cases, such as narrow earthen canals partially obscured by vegetation or shadows. Prior to model training, a comprehensive preprocessing pipeline was implemented to enhance data quality and model robustness. This process included radiometric and light correction, removal of low-quality and noisy images, normalization of pixel values, and cropping of images into standardized input sizes. The prepared dataset was subsequently divided into training and validation subsets. To mitigate overfitting and improve generalization, data augmentation techniques—such as rotation, flipping, scaling, and brightness variation—were applied exclusively to the training data. Six deep learning architectures were independently trained and evaluated for the task of canal segmentation. These networks represent a diverse range of encoder–decoder and lightweight segmentation designs commonly used in remote sensing applications. Each model was trained using the same dataset and evaluation protocol to ensure fair comparison. Hyperparameters, including learning rate, batch size, and optimizer configuration, were tuned individually for each network to achieve optimal performance. Model performance was assessed using standard semantic segmentation metrics, including Precision, Recall, Intersection over Union (IoU), and F1-score. These metrics provide complementary insights into detection accuracy, boundary delineation, and the balance between false positives and false negatives. In addition to quantitative evaluation, qualitative visual inspection of segmentation outputs was conducted to assess continuity, boundary accuracy, and performance in complex visual environments. Results The comparative evaluation of the six deep learning models reveals clear and consistent performance differences in irrigation canal segmentation. Among all tested architectures, Unet3+ demonstrated superior overall performance, achieving the highest Recall, IoU, and F1-score values. This indicates that Unet3+ is particularly effective at identifying both the presence and spatial extent of irrigation canals, including narrow and visually ambiguous earthen canals. While LinkNet achieved the highest Precision (93.96%), its comparatively lower Recall highlights an important limitation: although it accurately classified many canal pixels, it failed to detect a significant portion of less visible or thinner canals. This finding underscores the importance of balanced evaluation metrics such as F1-score and IoU, especially for applications where omission errors can lead to incomplete infrastructure mapping. The strong performance of Unet3+ can be attributed to its architectural design, specifically its full-scale multi-path skip connections that transfer information from all encoder layers to the decoder. This design enables the network to effectively integrate low-level spatial details with high-level semantic features, enhancing its ability to detect subtle visual cues. As a result, Unet3+ performs particularly well in complex scenes where canals are spectrally similar to surrounding soil, partially covered by vegetation, or affected by shadows. Qualitative results further confirm these findings, with Unet3+ producing more continuous and geometrically accurate canal boundaries compared to other models. The network demonstrates robustness across different canal types and environmental conditions, reducing fragmentation and misclassification commonly observed in other architectures. Comparison with previous studies highlights a key contribution of this research. While models such as LinkNet have shown strong performance in segmenting well-defined natural water bodies, their effectiveness decreases when applied to engineered and degraded canal systems. This study demonstrates that irrigation canals—especially earthen canals—pose distinct challenges that require more advanced feature aggregation mechanisms, such as those provided by Unet3+. Conclusion This research presents a comprehensive evaluation of deep learning–based semantic segmentation models for identifying agricultural irrigation canals from satellite imagery. Using a carefully prepared dataset from Khuzestan Province and a standardized evaluation framework, the study demonstrates that Unet3+ outperforms other widely used architectures in terms of detection accuracy, boundary preservation, and robustness across complex visual conditions. The results confirm that Unet3+ is particularly well suited for mapping narrow and irregular earthen canals, which are often absent from formal maps and difficult to survey manually. Its balanced performance across Precision, Recall, F1-score, and IoU makes it a reliable tool for real-world applications where both completeness and accuracy are critical. From a practical perspective, the findings have significant implications for irrigation management, water resource monitoring, and precision agriculture. Automated and accurate canal mapping enables better assessment of irrigation efficiency, identification of water loss pathways, and informed decision-making for maintenance and infrastructure development. This is especially important in water-stressed regions, where optimizing irrigation networks can contribute directly to water conservation and agricultural sustainability. Despite its strengths, the study also identifies areas for future improvement. The exclusive reliance on RGB satellite imagery may limit robustness under varying seasonal and illumination conditions. Integrating multispectral data, time-series imagery, or LiDAR information could further enhance segmentation performance. Additionally, transfer learning and domain adaptation techniques offer promising avenues for extending the applicability of the trained models to other geographic regions with minimal retraining. In conclusion, this research demo