Segmentation-Aware Latent Diffusion for Satellite Image Super-Resolution: Enabling Smallholder Farm Boundary Delineation

Abstract

This paper addresses the challenge of accurately delineating smallholder farm boundaries from low-resolution satellite imagery using advanced super-resolution techniques. We introduce a novel segmentation-aware latent diffusion model designed to enhance satellite image resolution while preserving critical boundary information for agricultural land use. The model integrates explicit segmentation guidance during the diffusion process to improve feature preservation relevant to farm boundaries. Experimental results demonstrate that the proposed method significantly outperforms existing super-resolution techniques in terms of image quality and subsequent farm boundary delineation accuracy, providing a robust solution for smallholder agriculture mapping.

1. Introduction

Smallholder farms are crucial for global food security, yet their accurate mapping and monitoring are hindered by the inherent low resolution of widely available satellite imagery. Traditional super-resolution methods often struggle to preserve fine-grained structural details essential for precise farm boundary delineation, leading to inaccuracies in agricultural assessments. This work addresses the critical need for enhanced satellite imagery to facilitate robust smallholder farm mapping, a challenge exacerbated by diverse farm sizes and irregular shapes. Models used in the article include: Latent Diffusion Model (LDM), U-Net, and Generative Adversarial Networks (GANs).

2. Related Work

Previous research in satellite image super-resolution has explored various deep learning architectures, including convolutional neural networks (CNNs), generative adversarial networks (GANs), and more recently, standard diffusion models. While these methods have shown promise in general image enhancement, they often fall short in applications requiring precise object boundary preservation, such as agricultural field segmentation. Efforts to integrate segmentation tasks with super-resolution typically involve multi-task learning, yet few specifically address the unique challenges posed by smallholder farm boundary delineation in complex agricultural landscapes.

3. Methodology

We propose a Segmentation-Aware Latent Diffusion (SALD) framework that integrates semantic segmentation information directly into the super-resolution diffusion process. The methodology involves an encoder-decoder diffusion network operating in a latent space, conditioned by both the low-resolution input image and a coarsely predicted segmentation mask. A novel segmentation-preservation loss function is introduced, guiding the denoising process to prioritize the accuracy and sharpness of potential farm boundaries during high-resolution image generation. Training utilizes a diverse dataset of paired low-resolution satellite images, corresponding high-resolution ground truth, and expert-annotated farm boundary masks.

4. Experimental Results

Experimental evaluation on a custom dataset of smallholder farm regions demonstrated that the SALD model significantly improved both image quality and subsequent segmentation accuracy compared to state-of-the-art super-resolution baselines. Quantitatively, SALD achieved a PSNR of 31.5 dB and an SSIM of 0.89, outperforming methods like ESRGAN and SwinIR. More critically, when a separate segmentation model was applied to the super-resolved outputs, SALD-enhanced images led to a 12% increase in Intersection over Union (IoU) for farm boundary delineation. The table below summarizes these key performance metrics across different super-resolution approaches.

Method	PSNR (dB)	SSIM	IoU (Farm Boundaries)
Bicubic Interpolation	24.3	0.72	0.58
ESRGAN	29.8	0.85	0.69
SwinIR	30.5	0.87	0.73
SALD (Proposed)	31.5	0.89	0.82

The results table illustrates the superior performance of the Segmentation-Aware Latent Diffusion (SALD) model across standard image quality metrics (PSNR, SSIM) and, most importantly, on the downstream task of smallholder farm boundary delineation (IoU). Notably, SALD significantly improves the IoU metric, indicating its effectiveness in generating images that facilitate more accurate segmentation compared to other leading super-resolution techniques.

5. Discussion

The superior performance of the proposed Segmentation-Aware Latent Diffusion model highlights the critical advantage of incorporating task-specific guidance during the super-resolution process, especially for applications like precise boundary delineation. The improved IoU metrics directly translate to more accurate estimates of cultivated land area, enabling better agricultural planning, yield prediction, and policy formulation for smallholder farmers. While the model demonstrates robustness across various farm sizes, future work could explore its generalization to diverse crop types and satellite sensor data, potentially integrating temporal information for dynamic farm monitoring.