Extraction Methods For Farmland Spatial Information Based On High-resolution Digital Elevation Models

Accurately acquiring farmland spatial information is the foundation of many areas such as building digital farmland protection system,precise farmland flood routine simulations,and flood loss evaluation.With spatial data acquisition techniques developing,traditional ways can no longer meet the current level of demand in efficiency and precision.LiDAR-derived DEM,which is efficient and precise to acquire,has become one of the main approaches to obtain high-resolution DEM,and has been applied in fine digital terrain analysis in many cases.Normally,cultivated land parcels are not directly adjacent to each other.Instead,intervening agriculture facilities,which might include trails and ditches,can cover large areas.The extraction of such small objects means significant to measuring cultivated lands,calculating farmland's crop production potential and environmental carrying capacity,and collecting the distribution and structure of farmland water utilities.This paper aims to establish a method for extracting farmland spatial information based on high-resolution DEM,and to accurately extract trails,ditches,and fields both in plains and mountainous areas.To achieve this,paper divided farmland features into ribbon-like features(trails and ditches)and blocky features(fields),and researched on their extraction methods,respectively.Specific research contents are as follows.(1)An extraction method for trails and ditches was established,and the optimal values of key parameters were analyzed and tested.According to the basic characteristic that trails are convex and ditches are concave,a method for extracting "ditch candidates"and "trail candidates" was developed:the DEM surface was first smoothed to have only main terrains and big features left;and then the original surface and the smoothed surface were compared to detect convex and concave parts.According to the characteristic that trails and ditches are ribbon-like,post-process methods were developed for connecting ditch and trail polygons,and eliminating noise polygons.(2)A method for extracting "field candidates" based on section lines was developed.According to the characteristic that cultivated fields are flat as artificial terrains,it was analyzed that how section lines perform inside a field.By analyzing the forms on section lines in horizontal,vertical,NE-SW diagonal,and NW-SE diagonal direction,flat parts were detected.(3)A "field candidate" extraction method based on Maximum Entropy Model was developed.By cultivated field terrain's characteristics were discussed and analyzed,the Maximum Entropy Model was introduced into extracting field parcels.According to field parcels' inner characteristics,combining different terrain types,the selection of terrain factors was discussed.The calculation of Maximum Entropy Model,with selected terrain factors acted in as constraints,extracted "field candidates".(4)Methods for extracting farmland spatial information were experimented in different terrains.To verify the efficiency of accurately extracting trails,ditches,and fields in large-scale and multi-terrain farmlands,the above-mentioned methods were applied on an 88.8 km2 plain research area and a 150.0 km2 mountainous research area,respectively.Qualitative and quantitative accuracy assessments were made on the results,and the errors were analyzed and discussed.Through the experiments and the accuracy assessment,following conclusions were obtained:(1)the method for extracting trails and ditches performed well,especially in trail extraction,where the Kappa coefficients in both two research areas achieved 0.8.The method was proved to have the capacity of extracting one-pixel wide features and separating adhered-together features;(2)the field extraction in flat farmland achieved a quite high precision when using either of two methods;(3)different terrains led to different results:the extraction results in flat farmlands are generally more precise than those in mountainous terraces;(4)different methods got different results:the producer's accuracies obtained by the section line method achieved 95%while its user's accuracies were lower;on the contrary,the user's accuracies obtained by the maximum entropy method achieved 95%while its producer's accuracies were lower.The error sources were analyzed in-depth,which also offered a mind on future research directions.