Earth observation (EO) provides various multi-platform, multi-temporal, and multiresolutionremote sensing imagery for dynamic monitoring of planet Earth, witha wide variety of uses. Crop monitoring is a typical application, which involvestimely gathering of the information of crop types, boundaries, and dynamic changesduring the whole crop growth period. However, most of the existing datasets andbenchmarks focus on medium-resolution (≥ 10 m) classification of the main croptype by using satellite image time series (SITS), where the individual boundaries(parcels) and the dynamic changes of the crop cannot be obtained, due to thelimited spatial resolution and the lack of multi-season annotation. In this paper,a multi-platform, multi-temporal, and multi-resolution (M3) remote sensing cropsegmentation dataset (M3CropSeg) is introduced for very high resolution (VHR, 1m) crop semantic segmentation to instance segmentation and dynamic segmentation.Specifically, M3CropSeg contains 16311 pairs of airborne VHR (1 m) and SITS(10 m) images, with 45 crop types and 101k instance annotations, covering a26,000 km2 area of California in the U.S. M3CropSeg has various challenges,including M3 data fusion, class imbalance, fine-grained classification, and multilabelclassification. Three tracks are designed for M3CropSeg, i.e., M3 semanticsegmentation, M3 instance segmentation, and M3 dynamic segmentation, to obtainhigh-resolution pixel-level, parcel-level, and multi-season crop types, respectively.The corresponding benchmarks are also provided to address the above challenges,along with a variety of experimental analyses.

Earth observation (EO) provides various multi-platform, multi-temporal, and multiresolutionremote sensing imagery for dynamic monitoring of planet Earth, witha wide variety of uses. Crop monitoring is a typical application, which involvestimely gathering of the information of crop types, boundaries, and dynamic changesduring the whole crop growth period. However, most of the existing datasets andbenchmarks focus on medium-resolution (≥ 10 m) classification of the main croptype by using satellite image time series (SITS), where the individual boundaries(parcels) and the dynamic changes of the crop cannot be obtained, due to thelimited spatial resolution and the lack of multi-season annotation. In this paper,a multi-platform, multi-temporal, and multi-resolution (M3) remote sensing cropsegmentation dataset (M3CropSeg) is introduced for very high resolution (VHR, 1m) crop semantic segmentation to instance segmentation and dynamic segmentation.Specifically, M3CropSeg contains 16311 pairs of airborne VHR (1 m) and SITS(10 m) images, with 45 crop types and 101k instance annotations, covering a26,000 km2 area of California in the U.S. M3CropSeg has various challenges,including M3 data fusion, class imbalance, fine-grained classification, and multilabelclassification. Three tracks are designed for M3CropSeg, i.e., M3 semanticsegmentation, M3 instance segmentation, and M3 dynamic segmentation, to obtainhigh-resolution pixel-level, parcel-level, and multi-season crop types, respectively.The corresponding benchmarks are also provided to address the above challenges,along with a variety of experimental analyses.