I recently had Terrset and just started to learn it (mainly all components). I hope you'll have solution or suggestion on this. You can also send to Terrset support as well.

You can write matlab code or use mathwork.com

You could consider them in terms of variation of rainfall/population density, by operating a simple operation of map algebra and subtracting the oldest map from the most recent one.

However, if you have the same value for all the pixels in your study area, this layer will not affect the calibration of your MLP: you are basically feeding your network with an irrelevant feature, as this information will not describe the different land use change patterns you have in the study area.

If you are able to access to other socio-economic data, I suggest you to apply the methodology used in the following paper to create spatial explicit maps through which calibrate your model. The methodology is applied to another LULCC model (SLEUTH), but you can do the same with IDRISI: Article Modelling the impact of urban growth on agriculture and natu...

If you want to read more abouth how the MLP in IDRISI works, I suggest you to read Article Fuzzy definition of Rural Urban Interface: An application ba...

Hi Federico Amato,

Much thanks for the previous answer and the materials provided, the first article was really learningful.

I do understand that including a non spatially variable paramter will not do much in the calibration process. However my main concern is in the prediction of the amount of change.

See, classically, the approach would be: taking land cover at time points t0 and t1, calibrate the model, and then predict land cover at time t2 and validate the prediction against a reference map available at that time t2. In that case, let us consider one has population size at times t0, t1 and t2 (or rate of increase for 10-t1 and t1-t2 time intervals). Suppose that the rate of increase of population across those time points is not the same. My guess is, naturally, the land use demand during the validation phase (t1-t2) might be different than the one during calibration phase (t0-t1), under the drive of that different rate of increase.

So my real question would be... How can one take this into account ? Or is the reasonning flawed ?

Thanks

In my opinion you have to consider two points:

1) Calibrating what I would call the “transition potential” map (i.e. the map expressing for each cell the probability of being affected by a transition between two land covers). In IDRISI this is the output of the MPL. Subsequently, IDRISI uses a Markov Chain to define the change demand, i.e. the quantity of changes expected among each pairwise combination of land cover classes in a certain time.

As you said, to correctly calibrate the model, at least three land cover maps, corresponding to three different periods (i.e.T0,T1 and T2) are needed. The model is performed using the T0 and T1 land cover maps as input to generate a simulation at T2. Then you can compare the result with a reference map. To focus the evaluation of the calibration on MLP you should avoid in this stage the use of the Markovian definition of the change demand, using instead the real quantities of changes observed in the real maps during the period T1-T2. See paragraph 2.3 of Article Fuzzy definition of Rural Urban Interface: An application ba...

2) Socio-economic data. As far as I understood, your problem is that you only have one value for the whole study area. This means, as an example, one value for the population variation for the entire area you are considering. This feature is on the one hand influencing the quantity of changes, that is what you model through the Markov Chain in IDRISI. On the other, as you have a constant value for all the training pixels, this will not really affect the MLP, and therefore the transition potential. To overcome the issues, it would be extremely useful to have data at a lower scale, for example having the population variation in the period T1-T2 for each single municipality (or district) included in the study area. This would help the model in describing the different behaviors of land use changes over the region ypu are considering.

Dear Federico Amato Thanks for the previous response. Sorry I took time before getting back at this, as I was experimenting some ideas. Ultimately you pinpointed the "real problem" in my case. The area on which I am working is quite small (~40 km sq), as compared to lots of the application cases of LULC modelling I went through. As such, I cannot have population census per district, since my study area is quite a single district, strictly speaking, you see the point. I was actually thinking about a different approach. Land use change demand is calculated (in the Markovian approach) by observing transitions between two time points: t0 and t1 (classically), or (t1 and t2 as you are suggesting). Considering resulting values basically translate probabilities of transitionning from one land use class to another, can we not alter them somehow to account for population growth ? I mean, if we know for instance that population increase would lead to favouring some specific types of land use conversions (expansion of cultivated areas for food, for example), is there some approach I can use to alter only relevant probabilities when generating Markov transition probabilities matrices ? I would love to have some insight and thoughts on this.

In this case you mainly deal with the MC model, so either you autonomously define a "scenario" to force some transition probability to be higher than others, or you follow some algorithm modification of traditional MC. i would suggest the second approach; in this case probably a good reference is:

- Takada,T., Miyamoto,A., Hasegawa,S.F., 2010. Derivation of a yearly transition probability matrix for land-use dynamics and its applications. Landsc.Ecol.25,561–572

Hi

I have used a conceptual semi distributed model GR2M at the monthly time step, but it must be possible to do the same with the daily GR4J version

and I changed the soil fixed parameter value into a variable which values changes annually according to population growth impact on land use/land cover

We put that into a paper (Paturel et al., 2016, hydrological processes)

which is available here

Article Using land cover changes and demographic data to improve hyd...

Best

Gil

Dear Prof Gil Mahe,

Much thanks for the recommendation. I actually know well that paper, as well as the previous developpement made by Pierre Diello in his thesis work (2007) on Nakambe basin.

I actually got some interesting insights and ideas from this. However, in this work, we are highly interested in the mapping, modelling and forecasting of spatial distribution of LULC changes, we went in a different direction, that is LULC modelling.

Ultimately, the problem has been "solved". I believe I might share the approach we used in another writeup, hopefully soon.

Thanks again.

Hi

LULC modelling must include population dynamics one way or another isnt it?

Good discussion, Following

Yes indeed, Prof. Gil Mahe .Somehow, it has to, especially in Sahelian cropping systems where it has been showed in lots of previous studies that cultivated areas expansion is highly correlated with population growth.

The problem I was facing to, in my special case, is the MLP neural network I am working with tries to build empirical statistical relationship between drivers and observed transitions at the calibration stages. It accepts as input spatially distributed data. To forecast changes, traditionnal approaches uses a Markovian stochastic processes to predict land demand (see article below) which was not suitable enough for our purposes.

We ended up forecasting land use deman based on population growth. Steps used were trivial: we forecasted population comparing various growth models (exponential, logistic) and selected the most realistic one results. Then cultivated areas were regressed against population on a past period, then forecasted based on future population estimates. Some key assumptions were made for other LULCs considered, which helped deriving a transition probability matrix for future LULC distribution.

Article A Review of Models of Landscape Change

Thanks for the paper

Dear Srutisudha Mohanty

First and foremost, be aware that in your case, it is not the population count per se which will drive the model, but rather the difference in population count in different locations (that, population density). As such, some land cover classes will tend to grow/shrink were population count (density) is higher/lower, which is actually what you are looking for.

Second, you might have a unique population data value for your entire area, and you might want to redistribute it across space, while keeping the total sum of all the values equal to the initial population count given.

Look into the TerrSet Help System, a thorough description of the DISAGGREGATE module is given, along with the definition of input data, which should be prepared separately first before feeding them to the module.

To Roland

Do you have a paper that could illustrate your comment ?

Dear Prof. Gil Mahe,

We do have a submitted paper covering in great details how the process was implemented, which is still under review. I will be glad to keep you posted on the status of this paper as soon as it is accepted for publication.

Regards,

Roland.

Thanks Roland

Yes I am definitely interested

Best

Gil

Hi

Sorry I do kot know tersset

But i would say that it depends on what you want to do with LULC layer. If it is just for displaying you can use the most practical way for you.

If it is for using with other applications, then it depends on what requires this application.

Dear Soly Jayan

Yes, and the TerrSet documentation is very specific about that. The input map should be given in rectangular form, arranged in tiles of similar resolution (see https://clarklabs.org/wp-content/uploads/2016/10/Terrset-Manual.pdf).

However, within the TerrSet LCM software, you can easily define a raster mask corresponding to the outline of your custom region, and flag all pixels out of this region of space as non-relevant. Therefore, the modelling process will only be carried within the area of interest.

I have successfully applied this in this paper :

Article Spatial and transient modelling of land use/land cover (LULC...

Regards,

Roland.