Mapping and monitoring crops is a key step towards sustainable\nintensification of agriculture and addressing global food security. A dataset\nlike ImageNet that revolutionized computer vision applications can accelerate\ndevelopment of novel crop mapping techniques. Currently, the United States\nDepartment of Agriculture (USDA) annually releases the Cropland Data Layer\n(CDL) which contains crop labels at 30m resolution for the entire United States\nof America. While CDL is state of the art and is widely used for a number of\nagricultural applications, it has a number of limitations (e.g., pixelated\nerrors, labels carried over from previous errors and absence of input imagery\nalong with class labels). In this work, we create a new semantic segmentation\nbenchmark dataset, which we call CalCROP21, for the diverse crops in the\nCentral Valley region of California at 10m spatial resolution using a Google\nEarth Engine based robust image processing pipeline and a novel attention based\nspatio-temporal semantic segmentation algorithm STATT. STATT uses re-sampled\n(interpolated) CDL labels for training, but is able to generate a better\nprediction than CDL by leveraging spatial and temporal patterns in Sentinel2\nmulti-spectral image series to effectively capture phenologic differences\namongst crops and uses attention to reduce the impact of clouds and other\natmospheric disturbances. We also present a comprehensive evaluation to show\nthat STATT has significantly better results when compared to the resampled CDL\nlabels. We have released the dataset and the processing pipeline code for\ngenerating the benchmark dataset.\n