Robert C. Frohn

Satellite remote sensing of isolated wetlands using object-oriented classification of Landsat-7 data.

There has been an increasing interest in characterizing and mapping isolated depressional wetlands due to a 2001 U.S. Supreme Court decision that effectively removed their protected status. Our objective was to determine the utility of satellite remote sensing to accurately detect isolated wetlands. Image segmentation and object-oriented analysis were applied to Landsat-7 imagery from January and October 2000 to map isolated wetlands in the St. Johns River Water Management District of Alachua County, Florida. Accuracy for individual isolated wetlands was determined based on the intersection of reference and remotely sensed polygons. The January data yielded producer and user accuracies of 88% and 89%, respectively, for isolated wetlands larger than 0.5 acres (0.20 ha). Producer and user accuracies increased to 97% and 95%, respectively, for isolated wetlands larger than 2 acres (0.81 ha). Recently, the Federal Geographic Data Committee recommended that all U.S. wetlands 0.5 acres (0.20 ha) or larger should be mapped using 1-m aerial photography with an accuracy of 98%. That accuracy was nearly achieved in this study using a spatial resolution that is 900 times coarser. Satellite remote sensing provides an accurate, relatively inexpensive, and timely means for classifying isolated depressional wetlands on a regional or national basis.

Segmentation and object-oriented classification of wetlands in a karst Florida landscape using multi-season Landsat-7 ETM+ imagery

Segmentation and object-oriented processing of single-season and multi-season Landsat-7 Enhanced Thematic Mapper Plus (ETM+) data was utilized for the classification of wetlands in a 1560 km2 study area of north central Florida. This segmentation and object-oriented classification outperformed the traditional maximum likelihood algorithm (MLC) in accurately mapping wetlands, with overall accuracies of 90.2% (single-season imagery) and 90.8% (multi-season imagery), compared to overall accuracies for the MLC classifiers of 78.4 and 79.0%, respectively. Kappa coefficients were over 1.5-times greater for the segmentation/object-oriented classifications than for the MLC classifications, and producer and user accuracies were also higher. The producer accuracies of the segmentation/object-oriented classifications were 90.8% (single-season) and 91.6% (multi-season), compared to 70.6 and 74.4%, respectively, for the MLC classifications. User accuracies were 73.9 and 73.5% for the single-season and multi-season segmentation/object-oriented classifications, respectively, compared to 54.1% (single-season) and 55.0% (multi-season) for the MLC classifications. The use of multi-seasonal data resulted in only a slight increase in overall accuracy over the single-season imagery. This small increase was primarily due to better discrimination of riparian wetlands in the multi-season data. Segmentation and object-oriented processing provides a low-cost, high-accuracy method for classification of wetlands on a local, regional, or national basis.

Mapping isolated wetlands in a karst landscape: GIS and remote sensing methods.

GIS and remote sensing (RS) techniques using Landsat ETM+ and objectoriented segmentation were developed to identify depressional isolated wetlands in a >2,600 km2 mixed land use area of north-central Florida, USA. Both the standalone RS method and the combined GIS/RS method were successful at identifying isolated wetlands > 0.20 ha. Combining the GIS and RS methods yielded producer and user accuracies ranging from 93 to 100% and 86 to 95%, respectively. The methods utilized successfully mapped isolated wetlands and could be used to address questions surrounding national estimates and areal distribution of isolated wetlands.

Improving Artificial Neural Networks Using Texture Analysis and Decision Trees for the Classification of Land Cover

The purpose of this research was to improve artificial neural network (ANN) classification of land cover using texture analysis and decision trees. Three variants on ANN-based classifiers were applied to Landsat-7 data of southwestern Ohio for an Anderson Level-II land-cover classification: (1) the use of a customized architecture for each land-cover class; (2) the use of texture analysis for urban classes; and (3) the use of a decision tree (DT) classifier to refine the ANN output. An accuracy assessment was performed on the final ANN classification and compared to the USGS National Land Cover Data (NLCD). Overall accuracy of the ANN was 85%, compared to 69% for the NLCD. Producer accuracies for the ANN ranged from 64% to 96% compared to 29 to 86% for the NLCD. User accuracies for the ANN ranged from 71 to 99% compared to 36-87% for the NLCD. The ANN methodology will be used to classify the state of Ohio and could potentially be used on a national scale.

Multi-scale Image Segmentation and Object-Oriented Processing for Land Cover Classification

The purpose of this research is to evaluate the utility of image segmentation and object-oriented processing for land cover classification in Ohio. Eight level-II land cover categories were classified using a region-growing segmentation algorithm and object-oriented fuzzy classification membership functions. The overall accuracy of the classification was 93.6%. Producer accuracies ranged from 89.63% for urban/recreational grasses to 98.01% for water. User accuracies ranged from 90.83% for deciduous forest to 99.49% for water. The high classification accuracies are primarily due to: (1) the use of multiple scales in the segmentation process for classification of small to large phenomena at the appropriate scale; (2) integration of textural, contextual, shape, and spectral information in the classification process; and (3) use of multi-temporal data to capture both leaf-on and leaf-off properties of land cover categories.

Image-based logistic regression parameters of deforestation in Rondonia, Brazil

Achieves an improved understanding of those factors associated with deforestation in the tropics. We developed a logistic regression model of deforestation based on features detectable on Landsat imagery. The model was developed for a study area in Central Rondo/spl circ/nia, Brazil using images covering a 20-year time period. The model incorporates five main parameters in predicting the amount of deforestation on individual settlement lots. These include (1) proximity of a lot to nearest paved road (PPR); (2) proximity of a lot to nearest secondary road; (3) distance to nearest forest/cleared edge (PFC); (4) distance to a market center; (5) distance to a major urban center. PPR and PFC were the strongest indicators of deforestation. In Rondo/spl circ/nia, deforestation strongly follows the pattern of roads creating a fishbone network. Deforestation increases at the edges of forest clearings as agriculture becomes unproductive over time. Several other parameters including agriculture and pasture suitability indices and soil ranking, using ancillary GIS data were tested using logistic regression but accounted for little variability in the model. The model is compared to actual Landsat classifications and field verification of deforestation in cross validation scheme over the 20-year time period.

Satellite Mapping and Monitoring of Wild Rice

Wild rice is the number one natural commodity in the state of Minnesota, and a major food source for Native Americans. Wild rice was estimated and mapped with Landsat data from 1998-2004 on the Leech Lake Indian Reservation in Minnesota. Due to increased precipitation in late 1998-early 1999 the reservation had a 65.7% loss in wild-rice crop area in 1999. In 2000, the wild rice crop returned to 52% of the original area. By 2004, the crop area had increased to 66% of the original crop area before the crop loss. The rate of recovery of wild rice following crop loss should be reflected in Leech Lakes insurance policy for wild rice.

GeoTempo: a modular, end-to-end OGC sensor web

Environmental information is valuable for monitoring and remediation efforts only if the location and time of acquisition are recorded in parallel with other parameters such as temperature, pH, pCO2 etc. Similarly, environmental information must be searchable and retrievable on the basis of geolocation and time of acquisition as well as other parameters such as sensor type, measurement type, measurement value, units of measure and keyword among others. Here we describe one specific and successful strategy for the implementation of geospatially and temporally-enabled cyber-infrastructure focused on open source geospatial standards (Open Geospatial Consortium or OGC), network protocols, operating systems, database software and geospatial extensions. This paper describes a real-world OGC compliant sensor web for environmental monitoring and mapping of terrestrial and aquatic invasive species with geospatial and temporal query capabilities called GeoTempo as demonstrated in the Great Lakes region of the United States and Canada.

An evaluation of classification methods for level II land-cover categories in Ohio

The purpose of this research was to evaluate six classifiers applied to Landsat-7 data for accuracy of Level II land-cover categories in Ohio. These methods consist of (1) USGS National Land Cover Data; (2) the spectral angle mapper; (3) the maximum likelihood classifier; (4) the maximum likelihood classifier with texture analysis; (5) a recently introduced hybrid artificial neural network; (6) and a recently introduced modified image segmentation and object-oriented processing classifier. The segmentation object-oriented processing (SOOP) classifier outperformed all others with an overall accuracy of 93.8% and Kappa Coefficient of 0.93. SOOP was the only classifier to have by-class producer and user accuracies of 90% or higher for all land-cover categories. A modified artificial neural network (ANN) classifier had the second highest overall accuracy of 87.6% and Kappa of 0.85. The four remaining classifiers had overall accuracies less than 85%. The SOOP classifier was applied to Landsat-7 data to perform a level II land-cover classification for the state of Ohio.

Satellite remote sensing of wild rice

Wild rice (Zizania aquatica) is a primary staple for Native Americans in the northern United States and there is a strong need to timely map and monitor its production on American Indian Reservations. This paper describes a methodology for the detection and classification of wild rice using satellite remote sensing. Landsat-7 data were used to map and estimate wild rice crop areas for the Leech Lake American Indian Reservation in Minnesota. A 14-band dataset was created using bands 1-5 and 7 from Landsat-7 images for July and October 1999. Two additional bands of image texture were also created from the 15-metre panchromatic channel of Landsat-7 for each time period and added to the dataset. Since wild rice grows in standing water, a data transform was developed to identify wet vegetation in the dataset. The data transform was created from the multi-temporal dataset and based on a minimum noise fraction transformation, image textural analysis, and hierarchical data masking. A lake and water mask was created and applied to the image to isolate wild rice beds from the wet vegetation transform. Training data for wild rice were collected using a GPS mounted to a laptop computer containing the image data. A spectral matched filtering algorithm was then applied to the image transform using ground collected data to classify wild rice within the data mask. Final classified data were verified using wild rice data collected in the field and from aerial photographs. We estimated nearly 2000 hectares of wild rice available for harvest for 1999 within the boundaries of the Leech Lake American Indian Reservation. This is over a 65% decrease in crop area of wild rice from 1994 (5750 hectares). We were able to verify the 1999 crop loss and subsequent insurance claim for the Leech Lake American Indian Reservation. Such mapping with Landsat-7 provides a more accurate wild rice database than can be routinely updated with repeat Landsat-7 coverage.