Abstract: Limitations of single sensor and single cyanobacterial extraction method for long time series monitoring of cyanobacterial blooms in Taihu Lake, based on the multi-source image data of GF-1 and Landsat 8 from 2014 to 2023, the normalised vegetation index (NDVI) method, random forest (RF) method and random forest based on maximum interclass variance (Otsu-RF) method are used to extract cyanobacteria from Taihu Lake.Determination of the optimal cyanobacterial extraction method by comparative analysis the spatial and temporal characteristics of cyanobacterial blooms in Taihu Lake over the past ten years are revealed. The results show that:① The Otsu-RF method has the highest accuracy in extracting cyanobacterial blooms in different images, and it can better extract the sporadically distributed cyanobacteria; ② Compared with the GF-1 images, the texture of cyanobacterial pixels on the Landsat 8 fusion images is clearer, and the results of cyanobacterial bloom extraction are more accurate; ③ The intensity of cyanobacterial bloom outbreaks was higher in summer and fall and weaker in spring and winter in Taihu Lake from 2014 to 2023, of which the outbreaks were particularly severe in 2017 and 2020, with the annual average area of cyanobacteria in the whole area exceeding 300 km²;④ The cyanobacterial blooms of Taihu Lake in spring, summer, and fall were mostly found in the Zhushan Lake, Meiliang Lake, and the shore of the western Taihu Lake areas, and they occurred more often in the shore of the the southern Taihu Lake areas in winter.

Key words: Taihu Lake, cyanobacterial bloom, multi-source image, random forest (RF), Otsu, spatio temporal variation