Early Season Crop Type Mapping Using Multi-Temporal Remote Sensing Data

Generating accurate crop type mapping is the basement of the crop growth monitor and crop yield prediction.As multi-temporal remote sensing data could describe the phenological difference between the crops,image time series have been widely used to identify crop types.The drawbacks of crop classification are(1)"data missing" problems for the time series data,(2)lack of ground reference data and(3)the demand of identifying crop types at early-season.Therefore,we estimated the effect of image compositions for early-season crop classification,and also proposed new methods for early-season crop type mapping without the use of ground reference data.We derived the following conclusions based on our experiments:(1)We tried to used four image compositions(daily,8-day,16-day and 32-day)for early-season crop type identification and the results showed that the daily composition image time series generated the highest classification accuracy.16-day time series and 8-day time series have similar classification accuracy,which are lower than the daily composition but much higher than the 32-day image composition.As 10m-30m image cannot be acquired daily,this paper recommend to use the 16-day image time series for early-season crop type mapping.(2)We estimated the contrnbution of the features for crop identification and found that NDVI and NIR have high contribution.Then,we proposed the Improved Artificial Immune Network(IAIN)classifier based on the reference time series for crop classification,the new classifier could process the image time series with "missing value".Next,we tried to use the IAIN to identify crops at early-season and the result showed that the major crops in the study region could be identified 4-6 weeks earlier than the crop harvest,the classification accuracies were higher than 95%.(3)When the ground reference data cannot be acquired for training the classifier,this study proposed a new method to acquire training samples using the crop mapping result of the previous years and reference time series.When the image time series used for crop mapping is April-August,the training sample acquired using this method could identify crop types with accuracy higher than 95%.(4)As Landsat data and Sentinel-2 data could be combined to generate the 30m image time series with high temporal resolution,which make it possible to generate reference time series using 30m data.Therefore,we used Landsat and Sentinel-2 data in 2017 to generate reference time series and then use the reference time series to replace the ground training samples to identify crop types at early-season in 2018.Results showed that the major crops in the study could be identified 6～8 weeks before the harvest with accuracies higher than 85%.