Improving the prediction accuracy of working state of no-tillage seeder and cleaning the invalid data timely will improve the data quality and reduce the cache pressure of monitoring equipment. However, as the agricultural machinery moved back and forth in the farmland, monitoring equipment captured a large number of invalid images at both ends of the farmland or when the vehicle stopped. These images affected the accuracy of farmland operation quality evaluation and created congestion in transmission network. A data cleaning method based on multi-condition time series, mainly vehicle speed, seeding rate and instantaneous area, was proposed to deal with the periodic change of long time series of agricultural machinery in the farmland. The model included multiple long-short term memory (LSTM) and spatiotemporal feature channel fusion (CONCAT connect) to maintain the individual difference under multi-condition. The current time sequence state of the agricultural machinery working condition can be predicted, and the real-time cleaning state of the image capture system can be indirectly acquired. Due to screen and capture valid image from captured image every three minutes by cleaning state, the system achieved the maximum efficiency in transmission channel and memory space. The comparison results of different models after 40 iterations showed that the prediction accuracy of this method for both valid and invalid samples was over 85% and the average accuracy of image cleaning was 92.4%. The data cleaning results showed that about 63% of the redundant data was removed after data cleaning. Therefore, the research method took the working condition of no-tillage seeder as the basis of image cleaning was effective, which had high research value and application prospect.