Improving lake chlorophyll-a interpreting accuracy by combing spectral and texture features of remote sensing

Abstract

Cyanobacterial blooms in lakes fueled by increasing eutrophication have garnered global attention, and high-precision remote sensing retrieval of chlorophyll-a (Chla) is essential for monitoring the blooms. Previous studies have focused on the spectral features extracted from remote sensing images and their relationship with chlorophyll-a concentrations in water bodies, ignoring the texture features in remote sensing images which is beneficial to improve interpreting accuracy. This study explores the texture features in remote-sensing images. It proposes a retrieval method for estimating lake Chla concentration by combining spectral and texture features of remote sensing images. Remote sensing images from Landsat 5, and 8 were used to extract NIR-Red, GREEN-BLUE, MNDWI, and KIVU bands ratio. The gray-level co-occurrence matrix (GLCM) of remote sensing images was used to obtain a total of 8 texture features; then three texture indices were calculated using texture features. Finally, a random forest regression was used to establish a retrieval model of in-site Chla concentration from texture and spectral index. Results showed that texture features are significantly correlated with lake Chla concentration, and they can reflect the temporal and spatial distribution change of Chla. The retrieval model combining spectral and texture indices has better performance (R2 = 0.801, RMSE = 16.0 µg·L− 1) than the model without texture features(R2 = 0.746, RMSE = 16.2 µg·L− 1). The proposed model performance varies in different Chla concentration ranges and is excellent in predicting higher concentrations. This study evaluates the potential of incorporating texture features of remote sensing images in lake water quality estimation and provides a novel remote sensing method to better estimate lake Chla concentration.

Downscaling Sentinel-3 Chlorophyll-a Concentration for Inland Lakes Based on Multivariate Analysis and Gradient Boosting Decision Trees Regression

“…(4) Texture features, which include correlation, contrast, entropy and variance. Some researchers have suggested that texture information can enhance the precision of Chl-a concentration assessment [46], [47]. Thus, several texture features obtained by using the widely-used grey scale cooccurrence matrix (GLCM) method, were selected as Chl-a concentration factors.…”