KAN You See It? KANs and Sentinel for Effective and Explainable Crop Field Segmentation

TL;DR

This work tackles crop-field segmentation from Sentinel-1/2 imagery and the need for explainable models. It introduces U-KAN, which replaces the deepest U-Net layers with Kolmogorov-Arnold Network (KAN) components to enable learnable activations and improved efficiency, and compares it to a standard U-Net. Results show a 2 percentage point IoU improvement with roughly half the GFLOPs, and Grad-CAM analyses indicate that U-KAN emphasizes field boundaries and yields more plausible explanations; per-channel analysis also reveals that some spectral channels are less informative. The findings advance efficient, explainable semantic segmentation for agriculture and highlight opportunities for channel selection to further reduce computation while preserving accuracy.

Abstract

Segmentation of crop fields is essential for enhancing agricultural productivity, monitoring crop health, and promoting sustainable practices. Deep learning models adopted for this task must ensure accurate and reliable predictions to avoid economic losses and environmental impact. The newly proposed Kolmogorov-Arnold networks (KANs) offer promising advancements in the performance of neural networks. This paper analyzes the integration of KAN layers into the U-Net architecture (U-KAN) to segment crop fields using Sentinel-2 and Sentinel-1 satellite images and provides an analysis of the performance and explainability of these networks. Our findings indicate a 2\% improvement in IoU compared to the traditional full-convolutional U-Net model in fewer GFLOPs. Furthermore, gradient-based explanation techniques show that U-KAN predictions are highly plausible and that the network has a very high ability to focus on the boundaries of cultivated areas rather than on the areas themselves. The per-channel relevance analysis also reveals that some channels are irrelevant to this task.

Figures (6)

• Figure 1: Example image from the South Africa Crop Type dataset sa_crop_type_dataset, Sentinel-2. (a) displays the image from Sentinel-2 in RGB, and (b) shows the corresponding ground truth, with crop field areas for segmentation highlighted in yellow. (c) and (d) present the saliency maps generated by U-Net and U-KAN, respectively, where red pixels indicate the areas of highest network focus.
• Figure 2: U-Net architecture unet. It is characterized by a U-shape structure with a contracting path (left side) to capture context and a symmetric expanding path (right side) to enable precise localization.
• Figure 3: U-KAN ukan. It has the same U-shape as U-Net, but the deepest layers are implemented as Tok-KAN blocks. Tok-KAN blocks are composed of a Tokenization layer, a KAN layer, a Downsampling layer, and a Normalization layer.
• Figure 4: South Africa Crop Type dataset example. Sentinel-1 shows the grayscale of the VV polarization, while Sentinel-2 shows the RGB of the same area. The mask shows the cultivated area in white.
• Figure 5: Per-channel relevance examples of U-KAN. The figure shows the ground truth over the original RGB image (a) and the saliency map of all 12 channels (b). Images (c), (d), and (e) display saliency maps generated by obscuring channels corresponding to B01, B06, and B11, respectively.