Spatial methods for deriving crop rotation history

Abstract

Benefits of converting 11 years of remote sensing classification data into cropping history of agricultural fields included measuring lengths of rotation cycles and identifying specific sequences of intervening crops grown between final years of old grass seed stands and establishment of new ones. Spatial and non-spatial methods were complementary. Individualyear classification errors were often correctable in spreadsheet-based non-spatial analysis, whereas their presence in spatial data generally led to exclusion of fields from further analysis. Markov-model testing of non-spatial data revealed that year-to-year cropping sequences did not match average frequencies for transitions among crops grown in western Oregon, implying that rotations into new grass seed stands were influenced by growers' desires to achieve specific objectives. Moran's I spatial analysis of length of time between consecutive grass seed stands revealed that clustering of fields was relatively uncommon, with high and low value clusters only accounting for 7.1 and 6.2% of fields.

Agricultural Crop Change in the Willamette Valley, Oregon, from 2004 to 2017

“…The classification of the Landsat and other images contains 57 classification categories, which were grouped into four super classes on the basis of majority-rule of time from 2004 through 2014 [13]. The super classes, which were created to ensure the accuracy for general inquiries, were annually disturbed agriculture, perennial crops, forests, and urban development.…”

“…The validation data contain all the pixels not used in training, randomly subsampled down, when necessary, to match the size of the training data for each class. Several of the original 19 annually disturbed agricultural crops and 20 perennial crops were difficult to reliably separate from each other in published results from 2004 through 2014; as such, the three annual ryegrass land-use classes (2, 12, and 44) were consolidated into one revised class [1,12,13,18]. In the present analysis, we have also consolidated land-use classes 3, 7, 10, and 41 for similar reasons, providing a total of 35 agricultural land-use categories, comprising 16 annually disturbed crops and 19 perennial crops.…”

“…The previously reported 11-year classifications [12,13] used as many of the full 57 landuse categories were present in the ground-truth data for each August through July western Oregon cropping year for harvests from 2005 through 2014. Synthetic ground-truth data for 2004 was generated after initial ML remote sensing classifications were calculated for 2005 through 2011, primarily by employing four superclasses to define locations where cropping practices and other land-use categories were most stable and least likely to have changed between 2004 and 2005 [19].…”

“…The series of images were obtained by combining two datasets derived primarily from remote sensing classifications of Landsat images. The datasets, obtained from 2004 to 2014 in the first case, and through 2017 in the second case, have been used in several studies conducted by the USDA-Agricultural Research Service [1,12,13]. In the last two years, the classified images were further enhanced by additional sets of constraints, which eliminated the majority of situations where logically impossible crop sequences initially occurred.…”

“…The land-use classification with all 57 classes, the corresponding super-classes, and their description, as they were originally created by Mueller-Warrant [13], is presented in Tables 1 and 2. The abbreviations used to describe the classes are: PR = perennial ryegrass (Lolium perenne L.).…”