PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-feature combined cloud and cloud shadow detection in GaoFen-1 wide field of view imagery

Zhiwei Li, Huanfeng Shen, Huifang Li, Guisong Xia, Paolo Gamba, Liangpei Zhang

arXiv:1606.05415v4cs.CV

TL;DR

GF-1 WFV cloud and cloud-shadow detection is difficult because the imagery has limited spectral information and calibration constraints. The paper proposes an automatic multi-feature combined method using spectral, guided-filtering, geometric, and texture-based processing, and reports strong performance across GF-1 WFV imagery.

  • Problem

    GF-1 WFV cloud detection is difficult because insufficient spectral information, unfixed radiometric calibration parameters, and absent onboard calibration complicate accurate identification.

  • Method

    The MFC method refines cloud masks with guided filtering and combines spectral, geometric, and texture features for cloud and cloud-shadow detection.

  • Results

    MFC performs well across most land-cover types and accurately detects thin clouds and cloud shadows using only four optical bands.

  • Takeaways & Limitations

    MFC provides a cloud and cloud-shadow detection approach for GF-1 WFV imagery using its limited optical-band setting.

  • Takeaways & Limitations

    Multi-temporal cloud-detection methods require more scenes over a short period and relatively stable land cover.

Abstract

from arXiv · show

The wide field of view (WFV) imaging system onboard the Chinese GaoFen-1 (GF-1) optical satellite has a 16-m resolution and four-day revisit cycle for large-scale Earth observation. The advantages of the high temporal-spatial resolution and the wide field of view make the GF-1 WFV imagery very popular. However, cloud cover is an inevitable problem in GF-1 WFV imagery, which influences its precise application. Accurate cloud and cloud shadow detection in GF-1 WFV imagery is quite difficult due to the fact that there are only three visible bands and one near-infrared band. In this paper, an automatic multi-feature combined (MFC) method is proposed for cloud and cloud shadow detection in GF-1 WFV imagery. The MFC algorithm first implements threshold segmentation based on the spectral features and mask refinement based on guided filtering to generate a preliminary cloud mask. The geometric features are then used in combination with the texture features to improve the cloud detection results and produce the final cloud mask. Finally, the cloud shadow mask can be acquired by means of the cloud and shadow matching and follow-up correction process. The method was validated using 108 globally distributed scenes. The results indicate that MFC performs well under most conditions, and the average overall accuracy of MFC cloud detection is as high as 96.8%. In the contrastive analysis with the official provided cloud fractions, MFC shows a significant improvement in cloud fraction estimation, and achieves a high accuracy for the cloud and cloud shadow detection in the GF-1 WFV imagery with fewer spectral bands. The proposed method could be used as a preprocessing step in the future to monitor land-cover change, and it could also be easily extended to other optical satellite imagery which has a similar spectral setting.

1. Introduction

GF-1 WFV imagery supports large-scale, frequently revisited Earth observation, but cloud and cloud-shadow contamination limits its application and complicates accurate detection with sparse spectral information.

  • Why cloud detection matters: Clouds and shadows contaminate optical imagery, obscure or remove spectral information, and affect applications including classification, segmentation, and feature extraction.Cloud-cover estimation also supports imagery-availability evaluation.
  • Research gap: Existing cloud-removal and restoration methods can repair contaminated imagery but do not provide a specific automatic approach for extracting clouds.Accurate cloud and cloud-shadow extraction is presented as a way to reduce contamination effects on imagery applications.
  • GF-1 WFV imagery: GF-1 WFV combines 16-m spatial resolution, four-day temporal resolution, four multispectral bands, and an 800-km combined swath for large-scale observation.The bands span the visible to near-infrared regions.
  • Detection challenges: GF-1 WFV cloud detection is difficult because radiometric calibration parameters are unfixed, onboard calibration is absent, and thermal or absorption bands are unavailable.These constraints make accurate calibration and cloud identification difficult.
  • Detection challenges: Limited spectral information makes bright ground objects and thin clouds difficult to distinguish, while shadow screening and matching errors hinder complete cloud-shadow localization.The text identifies snow, buildings, coastlines, and varying underlying surfaces as complicating cases.
  • Research gap: The paper motivates combining additional geometric and texture features with spectral information to improve cloud and cloud-shadow detection under limited bands.This motivation follows the difficulty of separating clouds from bright surfaces and detecting thin clouds.

2. Background

Cloud and cloud-shadow detection research spans spectral, morphological, probabilistic, machine-learning, and multi-temporal approaches, while shadow detection commonly follows cloud detection. The paper proposes MFC for GF-1 WFV imagery, combining guided filtering with geometric and texture features and reporting accurate results with four optical bands.

  • Cloud detection approaches: Single-scene cloud detection methods reduce input-data requirements, whereas multi-temporal methods generally achieve higher accuracy but require repeated scenes with limited land-cover change.This makes multi-temporal approaches more suitable for relatively permanent land areas in high-temporal-resolution imagery.
  • Cloud detection approaches: Prior cloud-detection approaches include ACCA, Fmask, HOT, Markov random fields, morphological feature extraction, neural networks, and support vector machines.These methods target cloud or cloud-shadow detection across Landsat and other remote-sensing imagery.
  • Cloud detection approaches: HOT requires prior image knowledge to construct a spectral-space separation line, and it may overestimate clouds because it cannot adequately suppress land-surface information.The latter limitation is stated for HOT in the paper’s method discussion.
  • Cloud-shadow detection: Cloud-shadow detection usually follows cloud detection, and shadow location depends on cloud position, satellite viewing geometry, and solar angles.Geometric calculation can predict shadow location when cloud location and height plus sun and satellite positions are known.
  • MFC contribution: The proposed MFC method uses guided filtering for local mask refinement and combines geometric and texture features to reduce cloud and cloud-shadow commission errors.The method is designed for GF-1 WFV imagery.
  • MFC contribution: MFC performs well across most land-cover types and accurately detects thin clouds and cloud shadows using only four optical bands.These results support its use under the GF-1 WFV spectral setting.

3. The MFC algorithm

MFC combines spectral, guided-filtering, geometric, and texture features to detect clouds and cloud shadows in GF-1 WFV imagery. It progressively refines a rough cloud mask, removes bright non-cloud objects, and derives cloud shadows through matching and correction.

  • Initial spectral detection: MFC converts four-band TOA reflectance into a rough cloud mask using spectral threshold segmentation.The thresholds combine HOT, visible-band ratios, and red-band reflectance to capture core cloud regions.
  • Cloud-boundary refinement: Guided filtering refines the binary mask near boundaries, recovering thin clouds that spectral segmentation may miss.The filtered gray mask is converted back to a binary mask using a fixed threshold of 0.12.
  • Bright-object filtering: Geometric and texture features remove bright non-cloud objects from the refined cloud mask.The procedure evaluates merged objects individually, using object shape and texture information to distinguish clouds from bright surfaces.
  • Cloud-shadow detection: Cloud shadows are obtained through cloud-shadow matching followed by correction, with cloud labels taking priority when labels overlap.This produces the integrated cloud and cloud-shadow mask.
  • Design rationale: Spectral information alone cannot fully exclude bright buildings and water bodies while preserving thin clouds at cloud boundaries.This motivates the staged design: conservative initial detection, boundary refinement, and later filtering.

A. Geometric features

MFC uses area, fractal dimension, and length-to-width ratio to distinguish cloud objects from bright non-cloud surfaces. These metrics target shape complexity, elongation, and object size.

  • Object elongation: LWR represents the relationship between object length and width using the minimum enclosing rectangle.The rectangle dimensions correspond to the object’s length and width for calculating the ratio.
  • Geometric metrics: MFC uses area, FRAC, and LWR as the three geometric metrics for each candidate object.Area counts contained pixels, while FRAC and LWR characterize object shape and elongation.
  • Cloud-object discrimination: Cloud objects generally have relatively low FRAC and LWR values because their shapes are less complex and elongated.These characteristics help separate clouds from coastlines, roads, and buildings, which often have higher FRAC or LWR values.
  • Area constraint: Object area prevents large clouds with high FRAC or LWR values from being incorrectly excluded.The area criterion complements shape metrics when cloud objects are large or geometrically atypical.

B. Texture features

The method combines geometric and texture features to distinguish cloud objects from bright non-cloud objects in GF-1 WFV imagery. Local binary patterns characterize texture, while conservative object filtering reduces cloud commission errors.

  • Combined filtering: Geometric features first remove objects lacking cloud-like shapes, while texture features further distinguish remaining cloud and non-cloud objects.This combination addresses bright water bodies, snow, and patches whose shapes can resemble clouds.
  • Texture descriptor: LBP texture descriptors characterize local spatial structure by comparing each pixel with neighboring gray values.The implementation uses rotation-invariant LBP concepts and represents images with histograms of decimal codes.
  • Texture classification: MFC compares current-object LBP histograms with cloud and non-cloud templates using chi-square distances.Templates include two cloud classes and two non-cloud bright-object classes trained from manually selected samples.
  • Adaptive extraction: The texture window expands with object size to provide enough pixels for feature extraction without leaving residual pixels after exclusion.The filtering procedure is implemented conservatively to avoid mistakenly removing clouds.
  • Filtering outcome: Filtering clearly decreases cloud commission error, although some non-cloud objects remain in the final cloud masks.Examples show snow and bright water bodies being excluded from refined cloud masks, while some connected snow objects are retained.
  • Shadow extraction: Cloud shadows are extracted differently over land and water using near-infrared and mean visible-band reflectance, respectively.The fill-hole transformation identifies dark regions relative to surrounding pixels, with water-shadow ambiguity limiting the water-area procedure.

4. Experimental results

The MFC method was evaluated on 108 globally distributed GF-1 WFV scenes spanning multiple cameras, acquisition dates, and land-cover conditions. It achieved strong cloud detection and cloud-fraction estimation, while cloud-shadow detection was weaker and sensitive to low shadow prevalence and thin clouds.

  • Validation data: 108 globally distributed GF-1 WFV full scenes were used for quantitative validation across four cameras and diverse land-cover types.The images covered forest, barren, ice/snow, water, wetlands, and urban areas.
  • Reference data: Reference masks were manually delineated because ground observations of cloud and cloud shadow were generally unavailable.Manual boundary differences and drawing errors may introduce a small accuracy-assessment bias.
  • Cloud-fraction estimation: MAE decreased from 0.109 for the official method to 0.027 for MFC, while MRE decreased from 0.722 to 0.198.MFC cloud fractions were closer to reference masks than official cloud fractions.
  • Cloud-fraction estimation: MFC achieved R-square 0.951 and RMSE 5.25% against reference cloud cover, compared with 0.648 and 13.08% for the official method.The reported fit indicates improved cloud-fraction estimation relative to the official method.
  • Cloud detection: The average cloud overall accuracy was 96.80%, with average producer’s and user’s accuracies of 88.30% and 92.05%.More than 98% of validation images had cloud overall accuracy above 80%.
  • Performance by surface: MFC performed well in forest, grassland, barren, urban, and water areas, but snow/ice scenes had lower cloud overall accuracy.Visual examples also showed strong cloud detection alongside deficiencies in cloud-shadow detection.

5. Discussion

MFC offers precise and fast processing modes that trade computational cost against cloud-mask detail and accuracy. It performs well across most land-cover types but remains vulnerable to snow/ice, thin clouds, shadows, calibration variation, and wide viewing angles.

  • Computational efficiency: MFC takes less than 30 seconds for cloud-fraction estimation and 3–5 minutes to generate a precise cloud and cloud shadow mask for one GF-1 WFV scene.These timings were reported for scenes of about 17000×16000 pixels using a parallel C++ implementation on an Intel Core i5-4210M laptop.
  • Downsampling trade-offs: A subsampling ratio of 2 is recommended for precise mode, whereas a ratio of 6 balances processing time and accuracy for fast mode.The ratio of 2 improves processing speed with only a slight accuracy reduction; ratio 6 is recommended for rapid cloud-fraction estimation.
  • Processing modes: Fast-mode MFC rapidly estimates cloud fraction, while precise-mode MFC generates pixel-scale cloud and cloud shadow masks.Fast mode uses stronger downsampling and omits cloud shadow detection; precise mode preserves more detail and usually requires more time.
  • Residual errors: Thin clouds far from core cloud regions can be missed because refinement detects them only within core-cloud windows.Bright snow-covered areas and water bodies may also remain in cloud masks.
  • Cloud-shadow limitations: Cloud shadows may be confused with terrain shadows or nearby water bodies, and correction can increase commission errors when these objects remain in the potential-shadow layer.Large-area object filtering and low near-infrared reflectance contribute to shadow misclassification.
  • Calibration and viewing-angle effects: Radiometric calibration coefficients vary over time, causing scene-dependent cloud and water reflectance and possible detection errors.Large viewing angles may also require angle-specific thresholds because estimated cloud-shadow directions may not be accurate across an entire scene.

6. Conclusions

The conclusions identify limited spectral information as a central difficulty for GF-1 WFV cloud and cloud-shadow detection. MFC combines guided filtering with geometric and texture features, performs well across most conditions, and is proposed for preprocessing and extension to similarly equipped imagery.

  • Insufficient spectral information makes GF-1 WFV cloud and cloud-shadow detection difficult, especially for thin clouds and bright non-cloud objects.
  • Guided filtering captures thin clouds near boundaries and reduces cloud omission errors.
  • Geometric and texture features reduce commission errors by excluding non-cloud bright objects and non-shadow objects.
  • Multiple spatial features compensate for spectral deficiencies in GF-1 WFV cloud and cloud-shadow detection.
  • MFC performs well under most land-cover conditions, especially in vegetation-covered areas, and achieves high accuracy with limited spectral bands.
  • MFC provides better cloud-fraction estimation than the official method and is proposed for producing clear-sky images for land-cover change monitoring.

Geoscience and Remote Sensing, 54, 2682-2694.

This supplied section consists of bibliographic entries covering cloud detection, cloud-shadow detection, image filtering, texture analysis, and remote-sensing applications.

  • The references include studies on automated cloud and cloud-shadow detection in optical satellite imagery.
  • The bibliography includes work on image filtering, texture features, morphological analysis, and spatial modeling.
  • Several cited studies address cloud detection and related applications for Landsat, MODIS, AVHRR, GF-1, and other satellite data.
Loading 1606.05415v4…