Advancing irrigation uniformity monitoring through remote sensing: A deep-learning framework for identifying the source of non-uniformity

Efficient agricultural water management ensures crop productivity and sustainability amidst climate change and water scarcity. This study integrates remote sensing and deep learning to advance irrigation uniformity monitoring by identifying sources of non-uniformity. Sentinel-2 satellite imagery from 2021–2023 was processed to generate 159,088 NDVI images from 1382 center pivot irrigation systems in Mato Grosso, Brazil. These images were classified into nine categories: vegetated, not vegetated, emitters, mechanical problems, low pressure, management zones, operational, partial crop, and clouds. Artificial images mimicking these patterns pre-trained a DenseNet121 convolutional neural network (CNN), addressing the challenge of limited labeled training data. Fine-tuning with six subsets of satellite data (2000–20,000 images) enhanced performance, achieving a Hamming accuracy of 99 % and an Exact Match accuracy of 91 %. Class-specific metrics demonstrated high precision, recall, and F1 scores for most patterns, though underrepresented classes, like mechanical issues, showed lower performance. The methodology was applied to 80 pivots in Mato Grosso (January–October 2024) using 2752 images, integrating classification results with the Satellite-Derived Christiansen Uniformity Coefficient (SDCUC). Among the pivots, 45 showed high uniformity (>90 % SDCUC), with 10 exhibiting irrigation-related issues, and 28 facing non-irrigation challenges. Another 32 pivots had acceptable uniformity (80–90 %), with 9 linked to irrigation problems and 25 to non-irrigation issues. Finally, 3 pivots had low uniformity (<80 %), with all issues related to non-irrigation factors like partial crop coverage. This scalable approach offers actionable insights for addressing non-uniformity, improving irrigation efficiency, and supporting precision agriculture, large-scale water management, and policymaking.