Knowledge-Informed Deep Learning for Irrigation Type Mapping from Remote Sensing

Oishee Bintey Hoque, Nibir Chandra Mandal, Abhijin Adiga, Samarth Swarup, Sayjro Kossi Nouwakpo, Amanda Wilson, Madhav V. Marathe

Abstract

Accurate mapping of irrigation methods is crucial for sustainable agricultural practices and food systems. However, existing models that rely solely on spectral features from satellite imagery are ineffective due to the complexity of agricultural landscapes and limited training data, making this a challenging problem. We present Knowledge-Informed Irrigation Mapping (KIIM), a novel Swin-Transformer based approach that uses (i) a specialized projection matrix to encode crop to irrigation probability, (ii) a spatial attention map to identify agricultural lands from non-agricultural lands, (iii) bi-directional cross-attention to focus complementary information from different modalities, and (iv) a weighted ensemble for combining predictions from images and crop information. Our experimentation on five states in the US shows up to 22.9% (IoU) improvement over baseline with a 71.4% (IoU) improvement for hard-to-classify drip irrigation. In addition, we propose a two-phase transfer learning approach to enhance cross-state irrigation mapping, achieving a 51% IoU boost in a state with limited labeled data. The ability to achieve baseline performance with only 40% of the training data highlights its efficiency, reducing the dependency on extensive manual labeling efforts and making large-scale, automated irrigation mapping more feasible and cost-effective.

Introduction

Agriculture accounts for nearly 90% of global freshwater consumption. The specific method of irrigation—whether flood, sprinkler, or drip—significantly impacts water efficiency and environmental sustainability. However, mapping these methods at scale using satellite imagery is notoriously difficult. Standard vision models face three primary hurdles: spectral ambiguity, where different systems look identical; extreme class imbalance, where efficient methods like drip irrigation represent a tiny fraction of the data; and regional variance across different climates.

Comparison of satellite RGB imagery and the target irrigation classification mask.

Methods: The KIIM Framework

The Knowledge-Informed Irrigation Mapping (KIIM) framework moves beyond simple pixel-based classification by integrating domain-specific agricultural knowledge into a Swin-Transformer architecture.

The KIIM architecture integrates multi-spectral imagery with crop-specific knowledge through specialized attention and projection modules.

The architecture consists of four key modules:

  • Soft Attention Module: Utilizes a land mask to focus the model's attention on agricultural boundaries, filtering out non-farm features like roads or buildings.
  • Multi-Stream Module: Processes both RGB and vegetation indices (NDVI, NDWI, NDTI) using a bidirectional cross-attention mechanism to correlate moisture levels with plant health.
  • Projection Module: Encodes the "knowledge-informed" component by using a matrix that maps specific crop types to their likely irrigation methods (e.g., vineyards typically use drip).
  • Ensemble Module: A learnable system that balances image-derived features with crop-based probabilities to generate the final irrigation mask.

The model was trained using multi-source geospatial data from five U.S. states (Arizona, Utah, Washington, Colorado, and Florida), utilizing Landsat 8 imagery and USDA Cropland Data.

Results and Discussion

KIIM demonstrated significant performance gains over standard baselines like DeepLabV3+. Across the five states, the model achieved macro-IoU improvements between 10.3% and 22.9%.

A major breakthrough was observed in identifying drip irrigation. In Utah, where drip irrigation is extremely rare, KIIM improved the IoU for this class by 71.4% over the best baseline. Furthermore, the model's efficiency is noteworthy: it reached baseline performance levels using only 40% of the training data, suggesting high feasibility for large-scale deployment.

In data-scarce regions like Florida, a two-phase transfer learning approach—pre-training on a balanced multi-state dataset before local fine-tuning—yielded a 51% IoU boost, proving that irrigation patterns are transferable across different geographic regions.

Conclusion

KIIM addresses the critical challenges of spectral ambiguity and class imbalance in irrigation mapping. By combining satellite imagery with crop-specific knowledge, it provides a more accurate and efficient tool for water resource management. Future work will focus on adapting the framework to global landscapes where data availability and agricultural practices vary significantly, and ensuring the model remains robust as farming technologies evolve.

Summary updated

Abstract

Accurate mapping of irrigation methods is crucial for sustainable agricultural practices and food systems. However, existing models that rely solely on spectral features from satellite imagery are ineffective due to the complexity of agricultural landscapes and limited training data, making this a challenging problem. We present Knowledge-Informed Irrigation Mapping (KIIM), a novel Swin-Transformer based approach that uses (i) a specialized projection matrix to encode crop to irrigation probability, (ii) a spatial attention map to identify agricultural lands from non-agricultural lands, (iii) bi-directional cross-attention to focus complementary information from different modalities, and (iv) a weighted ensemble for combining predictions from images and crop information. Our experimentation on five states in the US shows up to 22.9\% (IoU) improvement over baseline with a 71.4% (IoU) improvement for hard-to-classify drip irrigation. In addition, we propose a two-phase transfer learning approach to enhance cross-state irrigation mapping, achieving a 51% IoU boost in a state with limited labeled data. The ability to achieve baseline performance with only 40% of the training data highlights its efficiency, reducing the dependency on extensive manual labeling efforts and making large-scale, automated irrigation mapping more feasible and cost-effective.
BibTeX
@inproceedings{hoque2025knowledge,
  author = {Oishee Bintey Hoque and Nibir Chandra Mandal and Abhijin Adiga and Samarth Swarup and Sayjro Kossi Nouwakpo and Amanda Wilson and Madhav V. Marathe},
  title = {{Knowledge-Informed Deep Learning for Irrigation Type Mapping from Remote Sensing}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1077},
}
Figure