# Enhanced Garlic Crop Identification Using Deep Learning Edge Detection and Multi-Source Feature Optimization with Random Forest

## TL;DR

This paper presents a new method for accurately identifying garlic crops using deep learning and optimized features, improving accuracy and reducing noise in agricultural mapping.

## Contribution

The novel integration of deep learning edge detection, multi-source feature optimization, and spatial constraints for garlic crop identification.

## Key findings

- Feature optimization increased overall accuracy from 0.91 to 0.93 and Kappa coefficient from 0.8654 to 0.8857.
- DexiNed achieved 94.16% F1-score for field boundary extraction.
- Spatial constraints reduced salt-and-pepper noise effectively in garlic crop mapping.

## Abstract

Garlic, as an important economic crop, plays a crucial role in the global agricultural production system. Accurate identification of garlic cultivation areas is of great significance for agricultural resource allocation and industrial development. Traditional crop identification methods face challenges of insufficient accuracy and spatial fragmentation in complex agricultural landscapes, limiting their effectiveness in precision agriculture applications. This study, focusing on Kaifeng City, Henan Province, developed an integrated technical framework for garlic identification that combines deep learning edge detection, multi-source feature optimization, and spatial constraint optimization. First, edge detection training samples were constructed using high-resolution Jilin-1 satellite data, and the DexiNed deep learning network was employed to achieve precise extraction of agricultural field boundaries. Second, Sentinel-1 SAR backscatter features, Sentinel-2 multispectral bands, and vegetation indices were integrated to construct a multi-dimensional feature space containing 28 candidate variables, with optimal feature subsets selected through random forest importance analysis combined with recursive feature elimination techniques. Finally, field boundaries were introduced as spatial constraints to optimize pixel-level classification results through majority voting, generating field-scale crop identification products. The results demonstrate that feature optimization improved overall accuracy from 0.91 to 0.93 and the Kappa coefficient from 0.8654 to 0.8857 by selecting 13 optimal features from 28 candidates. The DexiNed network achieved an F1-score of 94.16% for field boundary extraction. Spatial optimization using field constraints effectively eliminated salt-and-pepper noise, with successful validation in Kaifeng’s garlic.