Laserfiche WebLink
MEMO 93 <br />4. Classifications <br />Classification procedures used in historic lands mapping were adapted from 2001 classification methods <br />described in SPDSS Technical Memoranda 89.2, Section 2.5. The hierarchical classification procedure <br />began with the discrimination of irrigated lands from non-irrigated lands through interpretation and <br />analysis ofmulti-temporal NDVI data derived from Landsat imagery for 1987 and 1976. Next, using GIS <br />the irrigated lands classification was combined with DOQs to "back-date" 2001 irrigated parcel <br />boundaries to represent 1987 parcel boundaries. Parcel back-dating consisted of adding parcels, <br />modifying existing parcel shapes to represent historic year parcels apparent in satellite imagery, and <br />removing existing parcels that were not present during the historic year. Backdating of parcels was used <br />to modify 2001 parcels to 1987, and 1987 parcels to 1976. The parcel update process was changed <br />slightly for the 1976 to 1950s backdating where aerial photo indexes were used instead ofmulti-temporal <br />NDVI data. <br />Once the parcels were delineated, irrigated lands for 1987 and 1976 were classified into major crop types <br />using the ERDAS Imagine Maximum Likelihood Classifier (MLC'). The MLC' is a commonly used <br />method of classification in remote sensing, in which a pixel with the maximum likelihood is classified <br />into the corresponding class (e.g. Jensen, 1996, Lillesand and Kiefer, 2000). <br />4.1 Historic Irrigated Lands Classification <br />The classification of historic irrigated lands included two main processes: (1) developing and applying <br />NDVI thresholds and (2) irrigated parcel delineation using printed maps with parcels, and NDVI <br />threshold rasters. <br />4.1.1 NDVI Thresholding <br />Irrigated and non-irrigated crop type reference information was not available for historic mapping, thus <br />NDVI thresholds could not be calculated using reference data as they were for 2001 mapping as described <br />in SPDSS Technical Memoranda for Task 89.2. However, the NDVI threshold values for 2001 were used <br />to estimate a 2001 "relative" threshold value (2001 RTV) that was then used to estimate historical NDVI <br />threshold values. The 2001 RTV is computed from the mean and standard deviation for each 2001 NDVI <br />date of a given Landsat frame using the following equation: <br />2001 RTV = 2001 NDVI Standard Deviation <br />(2001 Threshold - 2001 NDVI Mean) <br />This 2001 RTV was then used to compute the 1987 threshold value with the following equation: <br />1987 Threshold = 1987 NDVI Standard Deviation + Mean 1987 NDVI <br />2001 RTV <br />The threshold estimation process was repeated to compute 1976 threshold values from 2001 thresholds. <br />Threshold values were applied as a mask to NDVI composite rasters for both 1987 and 1976 Landsat <br />i The Maximum Likelihood Classifier MLC is based on the probability that a pixel belongs to a parkicular class. The basic <br />equation assumes that these probabilities are equal for all classes, and that the input bands have normal distribution. The <br />maximum likelihood classifier uses the Gaussian threshold stored in each class signature to determine if a given pixel falls within <br />the class or not. The threshold is the radius (in standard deviation units) of a hyperellipse surrounding the mean of the class in <br />feature space. If the pixel falls inside the hyperellipse, it is assigned to the class. The class bias (BIAS) is used to resolve <br />overlap between classes, and weights one class in favor of another. <br />Page 9 of 59 tij,Rarera~de ~eshn~f~gy, ~os~ <br />