Feature extraction and machine learning for the classification of active cropland in the Aral Sea Basin

Agricultural monitoring over large areas such as the irrigation zones of the Aral Sea Basin (ASB) needs Remote Sensing imagery with a high temporal resolution and a large swath width. Satellite sensors such as MODIS or the upcoming Sentinel-3 suits these requirements. A clear discrimination of active cropland coverage however is impossible using low spatial resolution imagery due to mixed pixel effects with other landcover types. In this paper we utilized Landsat 8 data and present the derivation of the active cropland in the ASB for 3 consecutive years (2014-2016). Each observed year comprises two growing seasons (spring and summer) with 40 Landsat scenes in each season, with a total of 240 Landsat scenes. Due to the lack of adequate in-situ data for remote agricultural regions, a feature extraction method based on image segmentation and k-means clustering is applied to delineate samples of agricultural fields which are forwarded to machine learning algorithms for image classification including. A simple markovian logic is used to introduce prior distribution values the following observation year. The results clearly show, that the extraction of active cropland is possible without the need for time series data, which is limited due to clouds and acquisition date irregularity. Model transferability tests were performed to assess machine learning inference on unseen data with training datasets from adjacent or non-adjacent scenes. The extraction of active cropland for each season also yielded secondary information such as the assessment of land use intensities and crop rotation patterns.

Hali tarjima qilinmagan