Estimating wheat fractional vegetation cover using a density peak k-means algorithm based on hyperspectral image data

• Main Authors: Da-zhong LIU, Fei-fei YANG, Sheng-ping LIU
• Format: Article
• Language: English
• Published: Elsevier 2021-11-01
• Series: Journal of Integrative Agriculture
• Subjects: fractional vegetation cover; k-means algorithm; NDVI; vegetation index; wheat
• Online Access: http://www.sciencedirect.com/science/article/pii/S2095311920635560

Fractional vegetation cover (FVC) is an important parameter to measure crop growth. In studies of crop growth monitoring, it is very important to extract FVC quickly and accurately. As the most widely used FVC extraction method, the photographic method has the advantages of simple operation and high extraction accuracy. However, when soil moisture and acquisition times vary, the extraction results are less accurate. To accommodate various conditions of FVC extraction, this study proposes a new FVC extraction method that extracts FVC from a normalized difference vegetation index (NDVI) greyscale image of wheat by using a density peak k-means (DPK-means) algorithm. In this study, Yangfumai 4 (YF4) planted in pots and Yangmai 16 (Y16) planted in the field were used as the research materials. With a hyperspectral imaging camera mounted on a tripod, ground hyperspectral images of winter wheat under different soil conditions (dry and wet) were collected at 1 m above the potted wheat canopy. Unmanned aerial vehicle (UAV) hyperspectral images of winter wheat at various stages were collected at 50 m above the field wheat canopy by a UAV equipped with a hyperspectral camera. The pixel dichotomy method and DPK-means algorithm were used to classify vegetation pixels and non-vegetation pixels in NDVI greyscale images of wheat, and the extraction effects of the two methods were compared and analysed. The results showed that extraction by pixel dichotomy was influenced by the acquisition conditions and its error distribution was relatively scattered, while the extraction effect of the DPK-means algorithm was less affected by the acquisition conditions and its error distribution was concentrated. The absolute values of error were 0.042 and 0.044, the root mean square errors (RMSE) were 0.028 and 0.030, and the fitting accuracy R2 of the FVC was 0.87 and 0.93, under dry and wet soil conditions and under various time conditions, respectively. This study found that the DPK-means algorithm was capable of achieving more accurate results than the pixel dichotomy method in various soil and time conditions and was an accurate and robust method for FVC extraction.