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

Introduction

Efficient water management is a cornerstone of global food security. While remote sensing has made significant strides in identifying irrigated versus non-irrigated land, distinguishing between specific irrigation methods—flood, sprinkler, and drip—remains a significant challenge. This paper introduces KIIM (Knowledge-Informed Irrigation Mapping), a framework designed to overcome spectral ambiguity and resolution constraints in satellite imagery by integrating agricultural domain knowledge.

The Challenge of Irrigation Mapping

Distinguishing irrigation types at a 30-meter resolution (e.g., Landsat 8) is difficult because drip infrastructure is often buried or too small to be seen. Furthermore, flood and sprinkler signatures can be spectrally similar at a distance. This technical difficulty is compounded by extreme class imbalance; in regions like Utah, drip irrigation accounts for only 0.08% of irrigated land, making it a "needle in a haystack" for standard machine learning models.

Comparison of RGB imagery and ground truth irrigation masks for various crops, highlighting the difficulty of visual identification.

Methodology: The KIIM Framework

KIIM moves beyond simple pixel-based classification by incorporating agricultural "common sense" into a multi-stream deep learning architecture. Key innovations include:

  • Multi-Stream Fusion with Cross-Attention: The model processes RGB imagery and Vegetation Indices (NDVI, NDWI, NDTI) through separate streams. A bidirectional cross-attention module allows these streams to prioritize the most relevant features from each modality, preventing "modality collapse."
  • The Projection Module: This is the core "knowledge-informed" component. It utilizes a probability matrix based on historical data that maps specific crop types to their likely irrigation methods (e.g., vineyards are highly likely to use drip). This acts as a powerful prior for the model when spectral data is ambiguous.
  • Two-Phase Transfer Learning: To handle regions with sparse labeled data, the model first learns universal irrigation patterns (like the circular shapes of center-pivot sprinklers) before fine-tuning on local regional nuances.
The KIIM architecture, featuring multi-stream fusion, soft attention, and the knowledge-informed projection module.

Key Results and Insights

KIIM demonstrates substantial improvements over state-of-the-art baselines like the Swin-Transformer:

  • Minority Class Accuracy: The model achieved a 71.4% IoU improvement for drip irrigation in Utah and 18.9% in Colorado.
  • Overall Performance: Across five U.S. states (AZ, CO, UT, WA, FL), the model saw macro-IoU improvements of up to 22.9%.
  • Data Efficiency: In Florida, the transfer learning approach allowed the model to reach baseline-level accuracy using only 40% of the typical training labels.

Conclusion

By bridging the gap between raw spectral data and agricultural context, KIIM provides a robust solution for large-scale irrigation mapping. The model proves that when physical infrastructure is below the sensor's resolution limit, crop-type context becomes the most effective proxy for accurate classification. Future work will focus on making the projection matrix dynamic to account for rapidly changing farming practices and expanding the model to international contexts.

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. Code: https://github.com/Nibir088/KIIM
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