Aiming at the problems of data quality degradation caused by multi-channel scanning-type loads on geostationary orbit remote sensing satellites in the process of imaging, transmission and storage, i.e., the influence of texture distortion and edge blurring in the meteorological remote sensing feature recognition images on the analysis of meteorological remote sensing images, this study proposes an improved BM3D noise reduction algorithm. The algorithm combines Morlet wavelet decomposition theory (with good symmetry and its decay characteristics follow the exponential law, it is able to match the mutation signals in the meteorological remote sensing images, thus realising signal denoising) and BM3D denoising principle (a non-local filtering algorithm that includes two parts: block matching and 3D collaborative filtering. Block matching involves grouping image blocks similar to a given reference block and composing them into a 3D array). Firstly, the image decomposes using wavelet transform to get four components. Secondly, the meteorological remote sensing image decomposes into three levels with a total of ten components. Finally, each component denoises using a separate BM3D filter, and the output image of the 10 components reconstructs. The output reconstructed image views as an estimate of the desired image, capable of suppressing meteorological remote sensing image noise and preserving edge detail. Compared with the traditional BM3D denoising algorithm, the improved BM3D algorithm is able to reduce the computation by about one-fifth. The eight meteorological remote sensing images process by equalising the grayscale and adding additive Gaussian white noise with mean 0 and standard deviation σ and random impulse noise. The median filter (suitable for removing isolated noise such as pepper noise), mean filter (suitable for removing noise from images), NL-Bayes (suitable for smoothing images and preserving image details), BM3D algorithm and the improved BM3D algorithm also compare to process the images respectively, and based on the results of peak signal-to-noise ratio (according to the definition of peak signal-to-noise ratio, it considers as the main metric to evaluate the quality of an image and utilises to measure the degree of realism of an image, with higher values indicating better denoising effects) of the meteorological remote sensing images, it finds that the average PSNR gain of the algorithms proposed in this study is in the range of 0.39 dB to 4.45 dB. The above experimental results of meteorological remote sensing images indicate that the improved BM3D algorithm works better, especially in the mixed noise denoising of Gaussian white noise and impulse noise.