Bat Habitat Suitability Model Evolution

These pages capture the progress made as the model structure is matured and more data becomes available to train and test the models. The biggest problem I face at this stage is that I do not have sufficient coverage of the Island to guarantee that the model is not bias free. Given the large number of presence samples I have for the local Parkhurst Forest area, I have employed a bias grid to reduce the presence data in line with the typical 3hr surveys from the pilot study work this year. So, here's the first evolution of the model for the Barbastelle and Pipistrelle bat species:

1st Model Evolution (Trial 007)

Notes:

• 1. All environment grids used with a regularisation factor of 2 (approximately 100m extent)

  2. Bias grid used to reduce contribution of the Parkhurst Forest area samples to the same weighting as 3 hours of survey effort

  3. Currently all models show training and testing bias due to the limited area covered on the Island, but despite this, the models look to predict habitat use in areas identified in the Isle of Wight Woodland Bat Survey commissioned by the Peoples' Trust for Endangered Species in 2003-7 by Ian Davidson-Watts.

  4. Interpretation of logistic bias: 0.8-1.0 very favourable conditions, 0.5-0.7 favourable, 0.3 -0.4 unfavourable, 0.0 to 0.2 very unfavourable. A logistic probability of 0.5 and above indicates likely use of habitat, anything below 0.3 indicates unlikely use of habitat.

  5. Next set of trials to be run after increasing the geographic range of the presence samples. 

2nd Model Evolution (Trial 008)

Notes:

• 1. Conditions as the first evolution except that additional presence samples from the 2006-7 Woodland Bat survey included.

  2. Next stage is to assess the individual grid contributions and look to see what tuning can be carried out

  3. Also note that the Pipistrellus pipistrellus HSM is based on 50% training and 50% testing due to the large number of samples in the dataset.

3rd Model Evolution (Trial 026)

This 3rd evolution of the model employs a more accurate habitat background, I found a few areas where the algorithm used to set the grid cell habitat had not always worked as intended and so have had to fix these by hand via the pixel editor in GRASS. In addition, the presence sample set has been extended to include the records held by the Local Record Centre here on the Isle of Wight. This has increased the volume of data available for both training and testing. 

The main noticeable difference is the fact that the models are generally a more defined and better fit, however some species are not well represented in the data (Pnat and Rfer) and so these models cannot be considered as statistically significant.

The basic grids now look to be working effectively, so will start to look at additional derived grids such as land gradient and slope direction.

4th Model Evolution (Trial 027)

The 4th model evolution includes two additional grids, a terrain roughness index (TRI) grid and a compass aspect grid. Both of these grids have been derived from the elevation grid using the DEM (Digital Elevation Model) tools in QGIS. I've only included a few examples from this run as the changes are quite subtle, and I've now got to a point where I will start produce tuned models for each species.