Sequoyah National Wildlife Refuge waterfowl land cover and habitat classification using SPOT-5 imagery

Developing effective habitat conservation and management strategies. The relationship between available habitat and waterfowl numbers obtained from aerial survey transects is not well studied. To determine these relationships, multispectral SPOT-5 satellite imagery acquired for Sequoyah National Wildlife Refuge close to the time of waterfowl surveys was used to map habitat conditions. Robust Random Forest classification trees were used to model 16 land cover types using 416 reference locations collected in the field or derived from aerial photos close to or during waterfowl survey dates. The normalized difference vegetation index (NDVI), normalized difference water index (NDWI) and a simple ratio (SR) of red and near infrared bands were used to enhance classification accuracy for key habitat areas and abundance of water. Terrain variables such as slope, solar illumination and cosine transformed aspect derived from a digital elevation model (DEM) were also used to enhance habitat classification. Random Forest (RF) models were also compared to support vector machines (SVM) and cforest (CF) conditional inference trees. We used error matrices and the Kappa agreement statistic (K) to compare model results from each classifier. Results indicated that a tuned RF classifier showed better performance (K=0.73) than SVM (K=0.65) and unbiased cforest trees (K=0.63). Overall class agreement between similar RF and cforest models, designed to reduce predictor variable selection bias, was also relatively low (K=0.47). A final tuned RF model was selected resulting in 75% accuracy overall and was used to map habitat types for the refuge and surrounding landscape. We found that elevation and minimum noise fraction (MNF) bands were the most important predictor variables. MNF bands can help to reduce the number of correlated variables entering into a classification model when a larger number of correlated spectral bands are used. Similar forest types such as riverine, bottomland hardwood, and floodplain forest showed the greatest misclassification error. Overall, the RF model and SPOT-5 leaf-off imagery generated accurate land cover data for assessing habitat conditions during waterfowl surveys..