Christine E. Blinn

On-the-Fly Massively Multitemporal Change Detection Using Statistical Quality Control Charts and Landsat Data

One challenge to implementing spectral change detection algorithms using multitemporal Landsat data is that key dates and periods are often missing from the record due to weather disturbances and lapses in continuous coverage. This paper presents a method that utilizes residuals from harmonic regression over years of Landsat data, in conjunction with statistical quality control charts, to signal subtle disturbances in vegetative cover. These charts are able to detect changes from both deforestation and subtler forest degradation and thinning. First, harmonic regression residuals are computed after fitting models to interannual training data. These residual time series are then subjected to Shewhart X-bar control charts and exponentially weighted moving average charts. The Shewhart X-bar charts are also utilized in the algorithm to generate a data-driven cloud filter, effectively removing clouds and cloud shadows on a location-specific basis. Disturbed pixels are indicated when the charts signal a deviation from data-driven control limits. The methods are applied to a collection of loblolly pine ( Pinus taeda) stands in Alabama, USA. The results are compared with stands for which known thinning has occurred at known times. The method yielded an overall accuracy of 85%, with the particular result that it provided afforestation/deforestation maps on a per-image basis, producing new maps with each successive incorporated image. These maps matched very well with observed changes in aerial photography over the test period. Accordingly, the method is highly recommended for on-the-fly change detection, for changes in both land use and land management within a given land use.

An Adaptive Noise-Filtering Algorithm for AVIRIS Data With Implications for Classification Accuracy

This paper describes a new algorithm used to adaptively filter a remote-sensing data set based on signal-to-noise ratios (SNRs) once the maximum noise fraction has been applied. This algorithm uses Hermite splines to calculate the approximate area underneath the SNR curve as a function of band number, and that area is used to place bands into ldquobinsrdquo with other bands having similar SNRs. A median filter with a variable-sized kernel is then applied to each band, with the same size kernel used for each band in a particular bin. The proposed adaptive filters are applied to a hyperspectral image generated by the airborne visible/infrared imaging spectrometer sensor, and results are given for the identification of three different pine species located within the study area. The adaptive-filtering scheme improves image quality as shown by estimated SNRs. Classification accuracies of three pine species improved by more than 10% in the study area as compared to that achieved by the same discriminant method without adaptive spatial filtering.

Automated Forest Area Estimation Using Iterative Guided Spectral Class Rejection

USDA Forest Service Forest Inventory and Analysis (FIA) forest area estimates were derived from 4 Landsat ETMimages in Virginia and Minnesota classified using an automated hybrid classifier known as Iterative Guided Spectral Class Rejection (IGSCR). Training data were collected using region- growing initiated at random points within each image. The classified images were spatially post-processed using five different techniques. Image accuracy was assessed using the center land-use of all available FIA plots and subsets contain- ing plots with 50, 75 and 100 percent homogeneity. Overall accuracy (81.9 to 95.4 percent) increased with homogeneity of validation plots and decreased with frag- mentation (estimated by percent edge; r 2 = 0.932). Filter- ing effects were not consistently significant at the 95 per- cent level; however, the 3 � 3 majority filter significantly improved the accuracy of the most fragmented image. The now-automated IGSCR is a suitable candidate for operational forest area estimation, with strong potential for use in other application areas.

Estimating leaf area index at multiple heights within the understorey component of Loblolly pine forests from airborne discrete-return lidar

ABSTRACT Airborne discrete-return light detection and ranging (lidar) can be used to estimate leaf area index (LAI) with relatively high accuracy. This capacity was explored with regard to assessing the capability of estimating LAI at different heights at the plot level, in the presence of understorey vegetation, within intensively managed Loblolly pine forest in North Carolina, USA. Field measurements utilized the LI-COR LAI-2200 plant canopy analyser for field-based estimates of effective LAI at three elevations within each plot; these were on the ground (0.0 m) and 1.0 m and 2.5 m above the ground within the various understorey heights and densities. A number of new and previously existing lidar metrics and indices were calculated from the distribution of return heights, which have been identified as potentially strong predictors of LAI. A bivariate and stepwise regression approach was then applied to create models for the estimation of LAI from lidar-derived height distribution metrics. The results show that specific logarithm transformed laser penetration indices calculated using a height threshold (e.g. the number of returns below 2.5 m ratioed against all returns) as close to field LAI measurement height (e.g. 2.5 m) was more effective than other lidar metrics. LAI can be estimated for each of the three measurement heights within the understorey component explaining 67 to 76% of the variance (root mean square error 0.42–0.57). The indices that produced the highest correlations and which were selected in stepwise regression analysis were calculated using all returns. The results indicate that LAI can be estimated accurately using lidar data in pine plantation forest over a variety of stand conditions.

Approximating prediction uncertainty for random forest regression models

Abstract Machine learning approaches such as random forest have increased for the spatial modeling and mapping of continuous variables. Random forest is a non-parametric ensemble approach, and unlike traditional regression approaches there is no direct quantification of prediction error. Understanding prediction uncertainty is important when using model-based continuous maps as inputs to other modeling applications such as fire modeling. Here we use a Monte Carlo approach to quantify prediction uncertainty for random forest regression models. We test the approach by simulating maps of dependent and independent variables with known characteristics and comparing actual errors with prediction errors. Our approach produced conservative prediction intervals across most of the range of predicted values. However, because the Monte Carlo approach was data driven, prediction intervals were either too wide or too narrow in sparse parts of the prediction distribution. Overall, our approach provides reasonable estimates of prediction uncertainty for random forest regression models.

Cloud-Sourcing: Using an Online Labor Force to Detect Clouds and Cloud Shadows in Landsat Images

We recruit an online labor force through Amazon.com’s Mechanical Turk platform to identify clouds and cloud shadows in Landsat satellite images. We find that a large group of workers can be mobilized quickly and relatively inexpensively. Our results indicate that workers’ accuracy is insensitive to wage, but deteriorates with the complexity of images and with time-on-task. In most instances, human interpretation of cloud impacted area using a majority rule was more accurate than an automated algorithm (Fmask) commonly used to identify clouds and cloud shadows. However, cirrus-impacted pixels were better identified by Fmask than by human interpreters. Crowd-sourced interpretation of cloud impacted pixels appears to be a promising means by which to augment or potentially validate fully automated algorithms.