Abstract: Objectives: Night-time light images have been widely used in human activity analysis, social economy estimation and other aspects. However, on night-time light images, farmland fire pixels caused by burning fields are easily confused with the urban light pixels, which interferes with the socioeconomic assessments using night-time light images. Based on the radiance time series characteristics of pixels, the farmland fire pixels on night-time light images can be identified through the random forest method. Methods: In this study, 10 continental countries in southern Africa were taken as the research area. Based on Black Marble product produced by NPP/VIIRS data, three characteristics of the time series of night-time light radiance were constructed, and the random forest classification method was used to divide the pixels into farmland fire pixels, stable light pixels, and black pixels. Results: In this study, stratified random sampling was used to manually test the classification accuracy, using time series of nighttime light data, high resolution satellite images and land cover data. Results showed that the overall accuracy of pixel classification was 91.2%, the average producer accuracy was 91.9%, and the average user accuracy was 91.0%. Among them, the producer accuracy and user accuracy of farmland fire pixel classification were 86.4% and 92.6%, respectively. Conclusions: The random forest classification method was used to classify the pixels on night-time light images into farmland fire pixels, stable light pixels and black pixels with high accuracy. The method proposed in this study can be used to filter out farmland fire pixels in Black Marble products, so as to improve the evaluation accuracy for African social economy using night-light images.