Print PageLand cover data are widely used in ecology to estimate land cover change. Whether change is due to land use change or not, those changes are a major component affecting ecological systems. Land cover data used as a proxy for habitat have proven especially valuable for predicting the occurrence of species and, by extension, species distributions. Remote sensing offers a fast process to obtain land cover images at a large scale. But resulting maps and inferences represent a final product produced by a complicated process that can introduce error at many points. As a result, most land cover data obtained using remote sensing are characterized by classification error. We modeled land cover dynamics using a fixed number of discrete habitat states. We present an approach to estimation that properly deals with state uncertainty producing unbiased estimates of habitat state transition probabilities. This approach requires ground truth data that inform classification parameters. We considered the case in which true and observed habitat states are observed for the same geographic units( e.g., pixels) but also the problem set by scaling-up: data informing classification being obtained at one scale (e.g., matrix of pixel accuracy) but true state and transition probabilities being estimated at a larger scale representing an aggregation from the smaller scale.
Results/Conclusions
In a first step, we tested the above approach on data simulated with classification error. In a second step, we analyzed data of habitat change of the southeast region of the USA. For both, we compared estimates of habitat transition probabilities between analyses based on true habitats and analysis where habitats observed with misclassification are implicitly assumed to represent truth. Results showed a strong bias on estimates of transitions probabilities when misclassification was not accounted for. Furthermore, scaling-up does not necessarily decrease the bias. It can even be the opposite: a high accuracy at a lower scale can still lead to a strong bias at a larger scale, especially if habitats exhibit a strong turnover. In this way, habitat dynamics of the southeast region of the USA appear distorted if classification errors are not accounted for: estimates of transition probabilities are biased towards a higher rate of habitat turnover, which could be interpreted as an instability of the system whereas it seems mostly to be an artifact caused by pixel misclassification.
