An Efficient Additive Kolmogorov-Arnold Transformer for Point-Level Maize Localization in Unmanned Aerial Vehicle Imagery

TL;DR

This work tackles point-level maize localization in ultra-high-resolution UAV imagery by introducing the Additive Kolmogorov-Arnold Transformer (AKT), which integrates Padé KAN (PKAN) modules and PKAN Additive Attention (PAA) to efficiently model small-object features and multiscale spatial dependencies. A new Point-based Maize Localization (PML) dataset with nearly 2,000 high-resolution images and ~501k annotations supports realistic field conditions and rigorous evaluation. AKT achieves up to a 4.2% improvement in average F1-score over state-of-the-art baselines, while reducing FLOPs by about 12.6% and increasing throughput by ~20.7%, and demonstrates sub-2 cm accuracy in interplant spacing estimation. The combination of Kolmogorov-Arnold representations with efficient additive attention provides a scalable framework for precision agriculture tasks, enabling accurate plant localization, counting, and spacing analysis on high-resolution UAV data.

Abstract

High-resolution UAV photogrammetry has become a key technology for precision agriculture, enabling centimeter-level crop monitoring and point-level plant localization. However, point-level maize localization in UAV imagery remains challenging due to (1) extremely small object-to-pixel ratios, typically less than 0.1%, (2) prohibitive computational costs of quadratic attention on ultra-high-resolution images larger than 3000 x 4000 pixels, and (3) agricultural scene-specific complexities such as sparse object distribution and environmental variability that are poorly handled by general-purpose vision models. To address these challenges, we propose the Additive Kolmogorov-Arnold Transformer (AKT), which replaces conventional multilayer perceptrons with Pade Kolmogorov-Arnold Network (PKAN) modules to enhance functional expressivity for small-object feature extraction, and introduces PKAN Additive Attention (PAA) to model multiscale spatial dependencies with reduced computational complexity. In addition, we present the Point-based Maize Localization (PML) dataset, consisting of 1,928 high-resolution UAV images with approximately 501,000 point annotations collected under real field conditions. Extensive experiments show that AKT achieves an average F1-score of 62.8%, outperforming state-of-the-art methods by 4.2%, while reducing FLOPs by 12.6% and improving inference throughput by 20.7%. For downstream tasks, AKT attains a mean absolute error of 7.1 in stand counting and a root mean square error of 1.95-1.97 cm in interplant spacing estimation. These results demonstrate that integrating Kolmogorov-Arnold representation theory with efficient attention mechanisms offers an effective framework for high-resolution agricultural remote sensing.

Figures (9)

• Figure 1: Various localization datasets are shown from left to right: JHU-CrowdJhu-crowd, MIO-TCDMIO-TCD, DOTADOTA, and the proposed PML dataset (detailed in Section\ref{['sec:dataset']}). Unlike conventional localization datasets, PML is specifically designed for real-world maize scenarios, providing precise annotations tailored to agricultural scenarios. As observed in the images, real-world data presents several challenges, including a small object-to-pixel ratio, highly variable field conditions, and significant background noise, all of which make accurate model localization more challenging and demand robust feature extraction and learning strategies.
• Figure 2: The specific location for the UAV imagery collection of real maize data, along with representative sample images and definitions of UAV flight paths and associated parameters, are detailed in Section \ref{['sec:dataset']}. (a)The specific site and the sample of the real-world maize fields. The red triangular marker indicates the actual collection site at the Arlington Research Station. (b) Diagram of Drone Data Collection Methods. The Unmanned Aerial Vehicle (UAV) follows a preplanned flight path to cover the entire experimental area, ensuring the completeness of data collection.
• Figure 3: The sample of the point-based maize localization dataset includes examples enhanced with various augmentation techniques. These techniques include brightness adjustment, grayscale transformation, the addition of Gaussian noise, rotation, and cutmix, among others. These augmentations aim to increase the diversity of the dataset, helping to improve model generalization and performance under varying real-world conditions.
• Figure 4: Overview of the proposed Additive Kolmogorov-Arnold Transformer (AKT) network. The AKT consists of multiple encoder and decoder layers. First, a CNN-based backbone is used to extract feature maps $F$ from the input image. Then, $F$ are added with positional embeddings, generating $F_{p}$, which is subsequently fed into the PKAN-Transformer encoder. The encoder captures contextual information, aggregates spatial details, and outputs the encoded features $F_{e}$. The encoded output features $F_{e}$, along with instance queries processed by the PAA, KAN layer, and other components, are passed through various attention modules and PKAN modules in the decoder. This process produces the decoder output $F_{d}$, which is used to predict the maize core.
• Figure 5: Comparison of various self-attention module variants: (a) represents the classical self-attention mechanism in ViT VIT; (b) shows the separable self-attention from MobileViTv2 Separableatt, which simplifies the feature representation by reducing a matrix to a vector; (c) illustrates the swift self-attention from SwiftFormer Swiftformer, employing additive attention to eliminate costly matrix multiplication operations; and (d) presents the proposed PKAN Additive Attention (PAA) module. Unlike previous attention mechanisms, PAA replaces the MLP with PKAN while maintaining accuracy. Furthermore, the key-value interaction leverages additive operations, effectively avoiding the quadratic computational complexity typical of traditional self-attention mechanisms.