IrrNet: Advancing Irrigation Mapping with Incremental Patch Size Training on Remote Sensing Imagery

TL;DR

The paper tackles irrigation mapping from Landsat imagery, where label scarcity and field heterogeneity hinder accurate discrimination of irrigation types. It introduces a progressive patch-size training paradigm with transfer learning across multispectral Landsat channels within an encoder-decoder segmentation framework, coupled with a hybrid BCE-Dice loss. Empirical results on the WRLU Utah dataset and supplementary USR Idaho data show a ~20% improvement over state-of-the-art methods and demonstrate the beneficial impact of additional spectral bands, though thermal channels contribute less in this context. The approach provides a practical baseline for long-term irrigation monitoring using historical Landsat data and offers a framework applicable to other remote-sensing irrigation tasks.

Abstract

Irrigation mapping plays a crucial role in effective water management, essential for preserving both water quality and quantity, and is key to mitigating the global issue of water scarcity. The complexity of agricultural fields, adorned with diverse irrigation practices, especially when multiple systems coexist in close quarters, poses a unique challenge. This complexity is further compounded by the nature of Landsat's remote sensing data, where each pixel is rich with densely packed information, complicating the task of accurate irrigation mapping. In this study, we introduce an innovative approach that employs a progressive training method, which strategically increases patch sizes throughout the training process, utilizing datasets from Landsat 5 and 7, labeled with the WRLU dataset for precise labeling. This initial focus allows the model to capture detailed features, progressively shifting to broader, more general features as the patch size enlarges. Remarkably, our method enhances the performance of existing state-of-the-art models by approximately 20%. Furthermore, our analysis delves into the significance of incorporating various spectral bands into the model, assessing their impact on performance. The findings reveal that additional bands are instrumental in enabling the model to discern finer details more effectively. This work sets a new standard for leveraging remote sensing imagery in irrigation mapping.

Figures (9)

• Figure 1: Representing the heterogeneity of irrigation types involves close proximity, forming diverse structures in the field characterized by variations in geometric shape, color, and patterns, despite exhibiting similarities.
• Figure 2: Irrigation type predictions by a U-Net with ResNet50 as backbone and by integrating our proposed method. The predictions made by U-Net struggles to focus on fine details where by integrating proposed method the model performs better. Few areas have been marked in the image to show the better prediction by the models (red bounding boxes represent wrong prediction).
• Figure 3: Illustration of the method proposed to integrate with the encoder-decoder based approach.
• Figure 4: Visualization of each band (Blue, Green, Red, Shortwave Infrared (SWIR) 1, SWIR 2, Near Infrared (NIR), Thermal in an input image. An area is highlighted with a bounding box to exemplify how each band contains distinct information within an image.
• Figure 5: After performing CLAHE histogram equalization on the normalized image.