Improved identification of cotton cultivated areas by applying instance-based transfer learning on the time series of MODIS NDVI

Cotton is an important cash crop and strategic material in the world, as the main source of natural and renewable fiber for textiles. Accurate and timely cotton distribution maps are crucial for monitoring and managing cotton cultivation system. Remotely sensed data have been widely used in cropland mapping, whereas relatively less attention has been paid specifically to cotton mapping partly due to the difficulty in obtaining the training samples over large regions. To resolve this issue, this study introduced the instance-based transfer learning to identify the cotton cultivated areas using the remotely sensed images. The annual time series of normalized difference vegetation index (NDVI) derived from Moderate Resolution Imaging Spectroradiometer (MODIS) images was adopted as input features. The random forest (RF) and long short-term memory (LSTM) algorithms integrated with the Transfer AdaBoost (TrAdaBoost) algorithm were adopted to generate the RF-based and LSTM-based TrAdaBoost approaches, respectively. The experiments were conducted in Arkansas State of the United States, where the cropland data layer (CDL) was available and utilized as a source of auxiliary data. The results showed that both the RF-based and LSTM-based TrAdaBoost performed better than the original RF and LSTM under the condition that the training samples in the target domain were limited. The advantages of transfer learning were much greater when the percentage of training samples from the source domain equals 80%. The two transfer learning approaches were then applied to identify the cotton cultivated areas in Uzbekistan. The cotton areas detected by MODIS images in 2018 were agreed well with the statistics at the sub-national level, with the R2 values of 0.57 and 0.64, respectively. These results demonstrate the potential of the RF-based and LSTM-based TrAdaBoost approaches in generating the cotton distribution maps when there are few samples in the target domain.