Search

Vector GIS represents shorelines as polylines that show the boundaries between land and water. This article compares two methods to measure how boundaries move among time points. The Area method co...

Validation of models of Land Use and Cover Change often involves comparing maps of simulated and reference change. The interpretation of differences between simulated and reference change depends on the characteristics of the process being studied. Our paper focuses on validation of models of one-way land change processes that spread in space. Our objective is to develop a method for validation of one-way land change models, such that the method provides objective information about the spatial distribution of errors. Using distance analysis on reference data, we build a baseline model for comparison with simulations. We then simultaneously compare the four maps of reference at initial time, reference at final time, simulation at final time, and baseline at final time. We also use Total Operating Characteristic curves and multiple-resolution map comparison. We illustrate the methods with a simulation of forest insect infestations. The methods give insights concerning the reference data and the spatial distribution of misses, hits, and false alarms with respect to initial points of infestations. The new methods reveal that the simulations underestimated change near initial points of spread. The spatial distribution of errors is a topic of land change models that deserves attention. For models of one-way, geographically-spreading processes, we recommend that validation should distinguish between near and far allocation errors with respect to initial points of spread.

The data employed in urban gain modeling classes are often imbalanced, negatively affecting the accuracy of traditional and standard data mining and machine learning models. This study presents a n...

Evaluation of land cover change (LCC) is commonly done at the pixel level; however, the model’s purpose may be relevant at a different grain size. Thus, the same model may be good for one purpose but inappropriate for another. For conservation applications, it is crucial to assess land change simulations at the grain relevant for the assessment of biodiversity impacts. Our objective is to evaluate land cover change scenarios in Bolivia, at the pixel-level and grain relevant to biodiversity, to inform LCC models for biodiversity assessments. We created six deforestation simulations that varied deforestation allocation based on forest management units (national, province, and municipality), ecoregions, and carbon stocks. We evaluated the simulations at the pixel level, and the objective’s relevant grain size through stratified error decomposition. We assessed biodiversity impacts by comparing the quantity of reference and simulated deforestation within species ranges. The spatial allocation of deforestation differed across simulations; however, their pixel-level error were similar. The province and municipality land change simulations had the lowest allocation errors at the relevant grain despite their large pixel-level errors, and they showed the lowest biodiversity errors. The province simulation provided the best balance identifying both affected species composition and the area of impact. This work presents evidence of the importance of incorporating information regarding the purpose of the simulation during model evaluation and selection. Error decomposition allowed ignoring irrelevant errors, translating into meaningful assessments of biodiversity impacts. As opposed to pixel-level metrics, stratified errors identified models that characterized biodiversity impacts best.

Intensity Analysis is a mathematical framework to express differences within a set of categories that exist at multiple time points. The original version of Intensity Analysis communicated land cha...

Some popular metrics to evaluate land change simulation models are misleading. Therefore, land change scientists have called for the development of methods to evaluate various aspects of modelling applications. This article answers the call by giving novel methods to compare three types of land change: 1) reference change during the calibration time interval, 2) simulation change during the validation time interval, and 3) reference change during the validation time interval. We compare these changes by using Intensity Analysis’ three levels and the Figure of Merit’s four components: Misses, Hits, Wrong Hits and False Alarms. We illustrate the concepts by applying a Cellular Automata – Markov land change model to a case study in northeast Hungary. We used reference maps of five land categories to calibrate the model during 2000–2006, then to validate the simulation during 2006–2012. Intensity Analysis’ time interval level shows that the simulation change and the reference change decelerated from 2000–2006 to 2006–2012. Intensity Analysis’ category level shows that the simulation losses were less than what a pure Markov chain would have dictated. Intensity Analysis’ transition level shows that the model’s Markov algorithm simulated correctly that the gain of Forest targeted Agriculture and Wetland. The Figure of Merit’s components reveals more allocation error than quantity error. Our collection of metrics show that more error derived from the Cellular Automata algorithm than from the Markov algorithm. We recommend that scientists use Intensity Analysis and the Figure of Merit’s components to reveal various fundamental aspects of modelling applications.

This article uses the GlobeLand30 maps of land cover to characterize the difference between years 2000 and 2010 in Asia. Methods of Intensity Analysis and Difference Components dissect the transition matrix for nine categories: Barren, Grass, Cultivated, Forest, Shrub, Water, Artificial, Wetland and Ice. Results show that Barren, Grass, Cultivated, and Forest each account for more than 21% of Asia at both 2000 and 2010, while transitions among those four categories account for more than half of the temporal difference. Nearly ten percent of Asia shows overall temporal difference, which is the sum of three components: quantity, exchange and shift. Quantity accounts for less than a quarter of the temporal difference, while exchange accounts for three quarters of the temporal difference. The largest quantity components at the category level are a net gain of Barren and net losses of Grass and Shrub. Shrub demonstrates the most intensive loss and gain relative to a category’s size. The largest and most intensive transitions to Barren are from Grass and Shrub. The largest and most intensive transition to Artificial is from Cultivated. Error information is not available for GlobeLand30 concerning 2000 or temporal change, but a confusion matrix is available for the global extent at 2010. This article applies methods to interpret the difference between two time points when a confusion matrix is available for only the latter time point. If the 2010 global confusion matrix reflects errors in Asia, then such errors could help to explain some of the gross gain of Barren and the counter-intuitive loss of Artificial. If the GlobeLand30 data indicate true change, then gross change in Asia is 4.4 times larger than net change.

This article gives methods to analyze the difference between two datasets that describe a mutual set of categories. The methods can analyze classification error between mapped and reference categories, or change between two time points. Previous work showed how to compute difference size as the sum of three components: Quantity, Exchange and Shift. These components exist for difference by category and for difference overall. These components can be challenging to compare when the categories’ differences vary by size. To address this challenge, this article introduces equations to compute a component’s intensity, which is the size of the component divided by the size of the difference. Component intensities facilitate comparison of each category with other categories and with difference overall. The case study illustrates how to use component intensities to characterize temporal change using remotely sensed data. Results show how an intensive Exchange component can signal possible confusion of two categories with each other. The literature shows that authors could benefit from interpretation of component intensities. Readers can perform the calculations for free by using the diffeR package in R or the PontiusMatrix spreadsheet available at www.clarku.edu/∼rpontius .

Several factors influence the performance of land change simulation models. One potentially important factor is land category aggregation, which reduces the number of categories while having the potential to reduce also the size of apparent land change in the data. Our article compares how four methods to aggregate Corine Land Cover categories influence the size of land changes in various spatial extents and consequently influence the performance of 114 Cellular Automata-Markov simulation model runs. We calculated the reference change during the calibration interval, the reference change during the validation interval and the simulation change during the validation interval, along with five metrics of simulation performance, Figure of Merit and its four components: Misses, Hits, Wrong Hits and False Alarms. The Corine Standard Level 1 category aggregation reduced change more than any of the other aggregation methods. The model runs that used the Corine Standard Level 1 aggregation method tended to return lower sizes of changing areas and lower values of Misses, Hits, Wrong Hits and False Alarms, where Hits are correctly simulated changes. The behavior-based aggregation method maintained the most change while using fewer categories compared to the other aggregation methods. We recommend an aggregation method that maintains the size of the reference change during the calibration and validation intervals while reducing the number of categories, so the model uses the largest size of change while using fewer than the original number of categories.

Object-based image analysis (OBIA) has gained widespread popularity for creating maps from remotely sensed data. Researchers routinely claim that OBIA procedures outperform pixel-based procedures; however, it is not immediately obvious how to evaluate the degree to which an OBIA map compares to reference information in a manner that accounts for the fact that the OBIA map consists of objects that vary in size and shape. Our study reviews 209 journal articles concerning OBIA published between 2003 and 2017. We focus on the three stages of accuracy assessment: (1) sampling design, (2) response design and (3) accuracy analysis. First, we report the literature’s overall characteristics concerning OBIA accuracy assessment. Simple random sampling was the most used method among probability sampling strategies, slightly more than stratified sampling. Office interpreted remotely sensed data was the dominant reference source. The literature reported accuracies ranging from 42% to 96%, with an average of 85%. A third of the articles failed to give sufficient information concerning accuracy methodology such as sampling scheme and sample size. We found few studies that focused specifically on the accuracy of the segmentation. Second, we identify a recent increase of OBIA articles in using per-polygon approaches compared to per-pixel approaches for accuracy assessment. We clarify the impacts of the per-pixel versus the per-polygon approaches respectively on sampling, response design and accuracy analysis. Our review defines the technical and methodological needs in the current per-polygon approaches, such as polygon-based sampling, analysis of mixed polygons, matching of mapped with reference polygons and assessment of segmentation accuracy. Our review summarizes and discusses the current issues in object-based accuracy assessment to provide guidance for improved accuracy assessments for OBIA.

Changsha has undergone speedy socio-economic development, rapid modification of industrial structure, and acceleration of urbanization, which has influenced land cover change during the most recent three decades. Policies have aimed to conserve total agricultural area, but it is not clear how successful these policies have been. Our purpose is to characterize and interpret spatiotemporal patterns of land change with respect to the policy to maintain agricultural area in Changsha, China. Maps at 1990, 2000, and 2010 show four land categories: Built, Forest, Crop and Other. We compute change components and apply Intensity Analysis to compare the land changes during two time intervals: 1990–2000 and 2000–2010. We also compare the central region to the peripheral region during 1990–2010. The maps show that Changsha’s land change accelerated from 1990–2000 to 2000–2010. Change was more intensive in the central region than in the peripheral region. Crop and Forest experienced net decreases while Built experienced net increase during both time intervals and in both regions. Built’s gain targeted Crop and avoided Forest during both time intervals and in both regions. The central region’s largest change component is quantity change, due to Built’s net gain. The peripheral region’s largest change component is exchange, due to simultaneous transitions from Forest to Crop and from Crop to Forest. According to these data, policies have not maintained the quantity of Crop, as the peripheral region has not gained Crop sufficiently to compensate for Crop’s loss from the central region.

We compare two popular methods to quantify temporal change among categories: 1) Intensity Analysis and 2) the land use dynamic degrees. We apply the methods to measure land change both outside and inside the coastal zone of Longhai, which is a typical county-level coastal city in Southeast China. The maps show eight categories at four time points: 1986, 1996, 2002, and 2010. Intensity Analysis shows graphically the size and the intensity of changes at three increasingly detailed hierarchical levels in a manner that facilitates interpretation. In contrast, the comprehensive land use dynamic degree (CLUDD) has no practical interpretation because the CLUDD is the sum of loss intensities of categories that have different sizes. The single land use dynamic degree is the annual net change as a percentage of the initial size of the category, which offers limited information because the single degree: 1) fails to reveal the sizes of the category’s loss and gain, 2) does not indicate how each category contributes to total change, 3) is sensitive to the category’s initial size, and 4) reveals neither the size of the category’s annual net change nor the size of the category’s initial size.

This article compares and assesses eight remotely sensed maps of Philippine forest cover in the year 2010. We examined eight Forest versus Non-Forest maps reclassified from eight land cover products: the Philippine Land Cover, the Climate Change Initiative (CCI) Land Cover, the Landsat Vegetation Continuous Fields (VCF), the MODIS VCF, the MODIS Land Cover Type product (MCD12Q1), the Global Tree Canopy Cover, the ALOS-PALSAR Forest/Non-Forest Map, and the GlobeLand30. The reference data consisted of 9852 randomly distributed sample points interpreted from Google Earth. We created methods to assess the maps and their combinations. Results show that the percentage of the Philippines covered by forest ranges among the maps from a low of 23% for the Philippine Land Cover to a high of 67% for GlobeLand30. Landsat VCF estimates 36% forest cover, which is closest to the 37% estimate based on the reference data. The eight maps plus the reference data agree unanimously on 30% of the sample points, of which 11% are attributable to forest and 19% to non-forest. The overall disagreement between the reference data and Philippine Land Cover is 21%, which is the least among the eight Forest versus Non-Forest maps. About half of the 9852 points have a nested structure such that the forest in a given dataset is a subset of the forest in the datasets that have more forest than the given dataset. The variation among the maps regarding forest quantity and allocation relates to the combined effects of the various definitions of forest and classification errors. Scientists and policy makers must consider these insights when producing future forest cover maps and when establishing benchmarks for forest cover monitoring.

The profession debates how to encode a categorical variable for input to machine learning algorithms, such as neural networks. A conventional approach is to convert a categorical variable into a collection of binary variables, which causes a burdensome number of correlated variables. TerrSet’s Land Change Modeler proposes encoding a categorical variable onto the continuous closed interval from 0 to 1 based on each category’s Population Evidence Likelihood (PEL) for input to the Multi-Layer Perceptron, which is a type of neural network. We designed examples to test the wisdom of these encodings. The results show that encoding a categorical variable based on each category’s Sample Empirical Probability (SEP) produces results similar to binary encoding and superior to PEL encoding. The Multi-Layer Perceptron’s sigmoidal smoothing function can cause PEL encoding to produce nonsensical results, while SEP encoding produces straightforward results. We reveal the encoding methods by illustrating how a dependent variable gains across an independent variable that has four categories. The results show that PEL can differ substantially from SEP in ways that have important implications for practical extrapolations. If users must encode a categorical variable for input to a neural network, then we recommend SEP encoding, because SEP efficiently produces outputs that make sense.

Scientists frequently describe temporal change among land categories by reporting the single land use dynamic degree and the comprehensive land use dynamic degree (CLUDD). The original intention of CLUDD was to compute the annual change percentage, which is the size of annual change expressed as a percentage of the size of the spatial extent. However, the literature’s mathematical descriptions of CLUDD have been unclear, thus readers have imagined various ways to compute CLUDD. Our manuscript clarifies the confusion and offers rules for mathematical notation so that authors can avoid future confusion. We examine the literature to see how authors have computed and interpreted the land use dynamic degrees. We illustrate deficiencies of one version of CLUDD. Then we propose equations for the components of annual change percentage. The literature shows three common misunderstandings. First, some authors add percentages of categories without accounting for the sizes of the categories. Second, other authors compute either double or half of the annual change percentage. Third, many authors interpret CLUDD as if CLUDD were the annual change percentage when they use a version of CLUDD that is not the annual change percentage. We recommend that the professional community use annual change percentage, its three components and Intensity Analysis to express temporal change among categories. Vague mathematical notation has created confusion concerning the land use dynamic degrees; therefore, we give rules for how to write mathematics clearly in a manner that applies to many professions.

Remote sensing, or Earth Observation (EO), is increasingly used to understand Earth system dynamics and create continuous and categorical maps of biophysical properties and land cover, especially based on recent advances in machine learning (ML). ML models typically require large, spatially explicit training datasets to make accurate predictions. Training data (TD) are typically generated by digitizing polygons on high spatial-resolution imagery, by collecting in situ data, or by using pre-existing datasets. TD are often assumed to accurately represent the truth, but in practice almost always have error, stemming from (1) sample design, and (2) sample collection errors. The latter is particularly relevant for image-interpreted TD, an increasingly commonly used method due to its practicality and the increasing training sample size requirements of modern ML algorithms. TD errors can cause substantial errors in the maps created using ML algorithms, which may impact map use and interpretation. Despite these potential errors and their real-world consequences for map-based decisions, TD error is often not accounted for or reported in EO research. Here we review the current practices for collecting and handling TD. We identify the sources of TD error, and illustrate their impacts using several case studies representing different EO applications (infrastructure mapping, global surface flux estimates, and agricultural monitoring), and provide guidelines for minimizing and accounting for TD errors. To harmonize terminology, we distinguish TD from three other classes of data that should be used to create and assess ML models: training reference data, used to assess the quality of TD during data generation; validation data, used to iteratively improve models; and map reference data, used only for final accuracy assessment. We focus primarily on TD, but our advice is generally applicable to all four classes, and we ground our review in established best practices for map accuracy assessment literature. EO researchers should start by determining the tolerable levels of map error and appropriate error metrics. Next, TD error should be minimized during sample design by choosing a representative spatio-temporal collection strategy, by using spatially and temporally relevant imagery and ancillary data sources during TD creation, and by selecting a set of legend definitions supported by the data. Furthermore, TD error can be minimized during the collection of individual samples by using consensus-based collection strategies, by directly comparing interpreted training observations against expert-generated training reference data to derive TD error metrics, and by providing image interpreters with thorough application-specific training. We strongly advise that TD error is incorporated in model outputs, either directly in bias and variance estimates or, at a minimum, by documenting the sources and implications of error. TD should be fully documented and made available via an open TD repository, allowing others to replicate and assess its use. To guide researchers in this process, we propose three tiers of TD error accounting standards. Finally, we advise researchers to clearly communicate the magnitude and impacts of TD error on map outputs, with specific consideration given to the likely map audience.

This paper couples a Forward Feature Selection algorithm with Random Forest (FFS-RF) to create a transition index map, which then guides the spatial allocation for the extrapolation of urban growth using a Cellular Automata model. We used Landsat imagery to generate land cover maps at the years 1998, 2008, and 2018 for the Tehran-Karaj Region (TKR) in Iran. The FFS-RF considered the independent variables of slope, altitude, and distances from urban, crop, greenery, barren, and roads. The FFS-RF revealed temporal non-stationary of drivers from 1998–2008 to 2008–2018. The FFS-RF detected that altitude and distance from greenery were the most important drivers of urban growth during 1998–2008, then distances from crop and barren were the most important drivers during 2008–2018. We used the Total Operating Characteristic to evaluate the transition index maps. Validation during 2008–2018 showed that FFS-RF produced a transition index map that had predictive power no better than an allocation of urban growth near existing urban. Simulation to 2060 extrapolated that Tehran, Karaj, and their adjacent cities will interconnect spatially to form a gigantic city-region.

We apply a method to evaluate the strength of the evidence for deviations from uniform land change in a coastal area, in the context of Intensity Analysis. The errors in the CORINE maps at 1990 and 2006 can influence the apparent change, but the errors are unknown because error assessment of the 1990 map has never been released, while the error of the 2006 map has been checked for only some countries. The 1990 and the 2006 maps of a coastal watershed in Portugal served as the data to compute the intensities of changes among eight categories. We evaluate the sizes and types of errors that could explain deviations from uniform intensities. Errors in 2.0% of the 2006 map can explain all apparent deviations from uniform gains. Errors in 1.5% of the 1990 map can explain all apparent deviations from uniform losses. Errors in less than 0.7% of the 1990 map can explain all apparent deviations from uniform transitions to each gaining category. We analyse the strength of the evidence for deviations from uniform intensities in light of historical processes of change. Historical processes can explain some transitions that the data show, while the hypothesised errors in the data are the explanation for other transitions that are not consistent with known processes. Inconsistent transitions are an indication of the misclassification errors that could propagate to other land cover change applications, as in the assessment of hydrological processes.

Land change in Kigali, Rwanda, is examined using Intensity Analysis, which measures the temporal stationarity of changes among categories. Maps for 1981, 2002 and 2014 were produced that show the land categories Built, Vegetated and Other, which is composed mainly of croplands and bare surfaces. Land change accelerated from the first time interval (1981–2002) to the second time interval (2002–2014), as increased human and economic activities drove land transformation. During the first interval, Vegetated showed net loss whereas Built showed net gain, in spite of a small transition directly from Vegetated to Built. During the second interval, Vegetated showed net gain whereas Built showed nearly equal amounts of gross loss and gross gain. The gain of Built targeted Other during both time intervals. A substantial portion of overall change during both time intervals consisted of simultaneous transitions from Vegetated to Other in some locations and from Other to Vegetated in other locations.

Uncovering the associations between land use and river water quality is useful for managing land-based pollution in the catchment-coast continuum. However, it is not clear how land use affects water quality in the context of simultaneous point source (PS) pollution. In this study, we develop a self-organizing map (SOM)-based approach to explore the relationship between land use and water quality in the Minjiang River Watershed, Southeast China. Water samples from 139 headwater sub-watersheds were associated with six land use categories, namely, Woodland, Agriculture, Orchard, Built-up, Unused land and Water. Sampling sites are delineated into six clusters based on six water quality parameters: ammonium-N, nitrate-N, total nitrogen, soluble reactive phosphate, total phosphate and potassium permanganate index. Local relationships between land use and water quality among four clusters that have sufficient sample sizes are further identified. There is no significant land use-water quality correlation in one of the four clusters (including 37 sub-watersheds). And the greater the PS pollution is, the less significant the land use-water quality correlations are in clusters. The results demonstrate how PS pollution weakens the land use-water quality correlation. Our method can help to determine whether non-point source or PS pollution exerts greater influence on the quality of the water coming from watershed.

Conventional methods to analyze a transition matrix do not offer in-depth signals concerning land changes. The land change community needs an effective approach to visualize both the size and intensity of land transitions while considering possible map errors. We propose a framework that integrates error analysis, intensity analysis, and difference components, and then uses the framework to analyze land change in Nanchang, the capital city of Jiangxi province, China. We used remotely sensed data for six categories at four time points: 1989, 2000, 2008, and 2016. We had a confusion matrix for only 2016, which estimated that the map of 2016 had a 12% error, while the temporal difference during 2008–2016 was 22% of the spatial extent. Our tools revealed suspected errors at other years by analyzing the patterns of temporal difference. For example, the largest component of temporal difference was exchange, which could indicate map errors. Our framework identified categories that gained during one time interval then lost during the subsequent time interval, which raised the suspicion of map error. This proposed framework facilitated visualization of the size and intensity of land transitions while illustrating possible map errors that the profession routinely ignores.

Comparison between two time points of the same categorical variable for the same study extent can reveal changes among categories over time, such as transitions among land categories. If many categories exist, then analysis can be difficult to interpret. Category aggregation is the procedure that combines two or more categories to create a single broader category. Aggregation can simplify interpretation, and can also influence the sizes and types of changes. Some classifications have an a priori hierarchy to facilitate aggregation, but an a priori aggregation might make researchers blind to important category dynamics. We created an algorithm to aggregate categories in a sequence of steps based on the categories’ behaviors in terms of gross losses and gross gains. The behavior-based algorithm aggregates net gaining categories with net gaining categories and aggregates net losing categories with net losing categories, but never aggregates a net gaining category with a net losing category. The behavior-based algorithm at each step in the sequence maintains net change and maximizes swap change. We present a case study where data from 2001 and 2006 for 64 land categories indicate change on 17% of the study extent. The behavior-based algorithm produces a set of 10 categories that maintains nearly the original amount of change. In contrast, an a priori aggregation produces 10 categories while reducing the change to 9%. We offer a free computer program to perform the behavior-based aggregation.

This article presents the clearest description to date concerning how to compute the hypothetical classification errors that could explain deviations from uniform land changes in the context of Intensity Analysis. Intensity Analysis is an accounting framework that analyses a square contingency table to measure how the sizes of the changes compare to the sizes of the categories. Our case study analyses maps of a portion of southern Nigeria at 1987 and 2002 that show the categories: cultivation, forest, settlement, and water. The data reveal that the changes are not uniformly proportional to the sizes of the categories in terms of categorical gains, categorical losses, and transitions. The methods in this article show that hypothetical error in 9% of the 2002 map could explain all the deviations from uniform categorical gains, hypothetical error in 11% of the 1987 map could explain all the deviations from uniform categorical losses, and hypothetical error in 6% of the 1987 map could explain all deviations f...