Analysis of landscape pattern on urban land use based on GF-5 hyperspectral data

Jingwen Yuan, Chen Wu, Bo Du, Liangpei Zhang, Shugen WangNational 2020 · DOI

Abstract

Rapid urbanization necessitates precise land-use monitoring to support sustainable urban planning. This study utilizes hyperspectral imagery from China’s Gaofen-5 (GF-5) satellite to analyze the urban landscape of seven core administrative districts in Wuhan's Jiangbei area. We introduce SSUN-CRF, an improved deep learning framework that combines a Spectral-Spatial Unified Network (SSUN) with Fully Connected Conditional Random Fields (FC-CRF) for high-precision classification. Our results demonstrate an overall accuracy of 90.48% and a Kappa coefficient of 0.8807. Landscape analysis using metrics such as Percentage of Landscape (PLAND) and Shannon’s Diversity Index (SHDI) reveals distinct spatial characteristics, including the prevalence of "urban villages" in Hanyang and the commercial density of Jianghan. Furthermore, spatial analysis indicates a clear expansion trend toward the southeast and southwest, providing critical decision-making support for the Yangtze River Economic Belt development.

Introduction

Monitoring urban land use is fundamental to managing the complexities of modern city growth. Traditional multispectral sensors often lack the spectral resolution required to differentiate between materials with similar reflectance properties, such as different types of roofing or pavement. Hyperspectral remote sensing, with its hundreds of narrow bands, offers a "spectral fingerprint" for precise material identification. However, the high dimensionality and spatial complexity of urban environments pose significant challenges for standard classification algorithms. This paper addresses these challenges by applying an advanced deep learning model to GF-5 hyperspectral data to quantify the landscape patterns of Wuhan, a major hub in the Yangtze River Economic Belt.

Methodology: The SSUN-CRF Framework

The proposed SSUN-CRF model integrates spectral and spatial information through a dual-branch architecture, followed by a refinement stage:

  1. Spectral Branch (LSTM): Long Short-Term Memory networks are employed to process the hyperspectral data as a sequence. This allows the model to capture the subtle spectral signatures across 330 bands (400–2500 nm), effectively overcoming the "curse of dimensionality."
  2. Spatial Branch (MSCNN): A Multiscale Convolutional Neural Network extracts spatial features—such as texture and shape—at various scales, ensuring that both large industrial blocks and small residential patches are recognized.
  3. Refinement (FC-CRF): To mitigate the "salt-and-pepper" noise common in pixel-based classification, a Fully Connected Conditional Random Field is applied. By considering the relationships between all pixels in the image, the FC-CRF smooths homogeneous regions and sharpens boundaries between distinct land-use types.

Data and Classification Results

The study utilized GF-5 data captured in October 2018 over Wuhan, featuring a 30-meter spatial resolution. Ground truth was established using 11 distinct land-use categories based on national standards (GB50137-2011), including commercial, industrial, and three tiers of residential land.

The SSUN-CRF model achieved a superior Overall Accuracy (OA) of 90.48%, significantly outperforming traditional methods like Support Vector Machines (46.72%) and the standalone SSUN model (89.04%). While the model excelled at identifying water bodies and industrial zones (precision > 90%), linear features like roads remained challenging due to the 30-meter resolution limit, where narrow infrastructure often occupies only a fraction of a pixel.

Landscape Pattern Analysis

Using Fragstats, we calculated seven landscape indices to interpret the urban structure and ecological health of Wuhan’s core districts.

Patch-Level Insights: Residential and Commercial Dynamics

  • Urban Villages: Analysis of the Percentage of Landscape (PLAND) revealed that Hanyang District has a notably higher index for "Type 3 Residential" land compared to other districts. This quantitatively confirms a significant presence of "villages in the city"—informal or low-quality housing clusters integrated into the urban fabric.
  • Commercial Centrality: Jianghan District exhibited high Patch Density (PD) but a low Aggregation Index (AI) for commercial land. This indicates that commercial zones are both numerous and widely distributed throughout the district, cementing its role as Wuhan’s primary commercial and trade hub.

Landscape-Level Insights: Diversity and Expansion

  • Ecological Stability: Shannon’s Diversity Index (SHDI) and Shannon’s Evenness Index (SHEI) suggest that land use in the Jiangbei core area is generally balanced. This diversity contributes to better landscape ecological stability and diversity.
  • Directional Sprawl: Moving window analysis of PD and SHDI reveals that Wuhan’s expansion is not uniform. The city is sprawling outward from the core, with the most pronounced growth occurring along the southeast and southwest axes of the main urban area.

Conclusion

The integration of GF-5 hyperspectral data with the SSUN-CRF algorithm provides a robust tool for urban land-use mapping. By identifying specific landscape patterns—such as the distribution of urban villages and commercial hubs—this research offers actionable insights for Wuhan’s urban planning and the broader development of the Yangtze River Economic Belt. Future work will focus on incorporating multi-temporal data to track these expansion trends over time and integrating higher-resolution data to improve the detection of linear urban infrastructure.

Summary updated

Abstract

æœ¬æ–‡ä»¥æ­¦æ±‰å¸‚æ±ŸåŒ—æ ¸å¿ƒåŒºåŸŸ7个行政区为例,以“高分五号”(GF-5)é«˜å ‰è°±é¥æ„Ÿå½±åƒä¸ºå®žéªŒæ•°æ®ï¼Œæå‡ºäº†ä¸€ç§åŸºäºŽæ·±åº¦å­¦ä¹ çš„ç©ºâ€”è°±ç»“åˆé«˜å ‰è°±å½±åƒåˆ†ç±»ç®—æ³•(SSUN)的改进算法SSUN-CRFï¼Œé€šè¿‡å ¨è¿žæŽ¥æ¡ä»¶éšæœºåœº(FC-CRF)对SSUN分类结果进行细化,并结合选取的7ä¸ªæ™¯è§‚æ ¼å±€æŒ‡æ•°åˆ†åˆ«ä»Žç±»åž‹æ–‘å—æ°´å¹³ä»¥åŠæ™¯è§‚æ°´å¹³åˆ†æž7ä¸ªè¡Œæ”¿åŒºæ™¯è§‚æ ¼å±€ç‰¹ç‚¹åŠæ­¦æ±‰å¸‚åŸŽå¸‚åŒ–è¶‹åŠ¿ã€‚ç»“æžœè¡¨æ˜Žï¼šï¼ˆ1)SSUN-CRF算法能够快速准确地提取城市用地类型,总体精度可达0.9048,Kappa系数可达0.8807;(2ï¼‰ä»Žæ–‘å—ç±»åž‹æ°´å¹³ä¸Šåˆ†æžï¼šæ ¹æ®å„è¡Œæ”¿åŒºä¸åŒç±»åˆ«æ™¯è§‚ç±»åž‹ç™¾åˆ†æ¯”(PLAND)å¯ä»¥çœ‹å‡ºï¼Œæ­¦æ±‰å¸‚ä¸»åŸŽåŒºä»¥å± ä½ç”¨åœ°ä¸ºä¸»ï¼Œæ±‰é˜³åŒº3ç±»å± ä½ç”¨åœ°PLANDæŒ‡æ•°é«˜äºŽå ¶ä»–ä¸»åŸŽåŒºï¼Œè¡¨æ˜Žæ±‰é˜³å­˜åœ¨è¾ƒå¤§â€œåŸŽä¸­æ‘â€çŽ°è±¡ï¼›æ±Ÿæ±‰åŒºå•†ä¸šç”¨åœ°æ–‘å—å¯†åº¦(PD)高且聚合度(AI)低,即商业用地多且分布均匀,表明江汉区是武汉的重要商贸区;(3ï¼‰ä»Žæ™¯è§‚å±‚æ¬¡æ°´å¹³ä¸Šåˆ†æžï¼šæ ¹æ®é¦™å†œå¤šæ ·æ€§(SHDI)、香农均匀度(SHEI)æŒ‡æ•°åˆ†æžå¯ä»¥çœ‹å‡ºï¼Œæ­¦æ±‰å¸‚æ±ŸåŒ—åŒºåŸŸå„ç”¨åœ°ç±»åž‹æ€»ä½“åˆ†å¸ƒå‡è¡¡ï¼Œæ™¯è§‚ç”Ÿæ€å¤šæ ·æ€§å’Œç¨³å®šæ€§è¾ƒå¥½ï¼›ä»Žæ–‘å—å¯†åº¦(PD)å’Œé¦™å†œå¤šæ ·æ€§(SHDI)ç©ºé—´å¸ƒå±€åˆ†æžå¯çŸ¥æ­¦æ±‰å¸‚çš„åŸŽå¸‚æ‰©å¼ å‘ˆçŽ°ä»¥ä¸»åŸŽåŒºä¸ºæ ¸å¿ƒå‘å¤–è”“å»¶è¶‹åŠ¿ï¼Œä¸”æ²¿ä¸»åŸŽåŒºä¸œå—ã€è¥¿å—æ–¹å‘æ‰©å¼ è¶‹åŠ¿æ˜Žæ˜¾ã€‚æœ¬æ–‡ç ”ç©¶æˆæžœå¯¹æ­¦æ±‰å¸‚æ€»ä½“å‘å±•è§„åˆ’ã€ä¼˜åŒ–åœŸåœ°åˆ©ç”¨åŠå¯æŒç»­å‘å±•å»ºè®¾å ·æœ‰é‡è¦çš„æŒ‡å¯¼æ„ä¹‰ï¼Œè¿›ä¸€æ­¥å¯ä¸ºå‘å±•é•¿æ±Ÿç»æµŽå¸¦ã€â€œé•¿æ±Ÿä¸»è½´â€æä¾›å†³ç­–å‚è€ƒã€‚
BibTeX
@inproceedings{yuan2020analysis,
  author = {Jingwen YUAN and Chen Wu and Bo Du and Liangpei Zhang and Shugen WANG},
  title = {{Analysis of landscape pattern on urban land use based on GF-5 hyperspectral data}},
  booktitle = {National},
  year = {2020},
  doi = {10.11834/jrs.20209252},
}
Figure