S. Brosse et al., Predicting fish distribution in a mesotrophic lake by hydroacoustic surveyand artificial neural networks, LIMN OCEAN, 44(5), 1999, pp. 1293-1303
The present work describes the development and validation of Artificial Neu
ral Networks (ANN) by comparison with classical and more advanced parametri
c and nonparametric statistical modeling methods such as Multiple Regressio
n (MR), Generalized Additive Models (GAM), and Alternating Conditional Expe
ctations (ACE) to estimate spatial distribution of fish in a mesotrophic re
servoir. The modeling approaches were developed and tested using 60 hydroac
oustic transects covering the whole lake. Each transect was divided into 10
0-m-long sections, constituting a total of 732 sampling units. For each of
them, the relationships between topology, chemical characteristics, and fis
h abundance were studied. The models had six independent topological (i.e.,
depth, distance from the bank, slope, and stratum) and chemical (i.e., tem
perature and dissolved oxygen) variables and one dependent output variable
(fish density, FD). The data matrix was divided into two parts. The first c
ontained units where FD was nonnil (i.e., 399 sampling units), and the seco
nd contained only cases without fish (i.e., 333 sampling units). Model trai
ning and testing procedures were run on the first submatrix after log(FD 1) transformation. As linear MR results were not satisfactory (r(2) = 0.42
in the training set, and r(2) = 0.51 in the testing set) compared with ANN
(r(2) = 0.81 in the training set, and r(2) = 0.77 in the testing set), we t
ried nonlinear transformations of the variables such as logarithmic, lowess
(for the GAM), and an optimal nonlinear transformation using the SAS Trans
reg procedure (for the ACE model), but the determination coefficients remai
ned clearly lower than those obtained using ANN (r(2) = 0.60 in the trainin
g set for ACE, and r(2) = 0.66 in the training set for GAM). The results of
a second test on the nil submatrix stressed that, compared with other stat
istical techniques, ANN and, to a certain extent, GAM models were able to c
learly define the potential FDs in samples where no fish were actually foun
d. The model showed, on the basis of the topological and chemical variables
taken into account, that the predicted potential FDs in the surface stratu
m are higher than in the underlying stratum. Finally, on the basis of the s
ensitivity analyses performed on the ANN and GAM results, we established re
lationships between FDs and the six environmental variables. Our results ex
hibit a clear summer habitat preferendum, the fish (predominantly roach) be
ing located mainly in the surface stratum, in the warm shallow littoral are
as. These observations led us to discuss the ecological significance of suc
h a fish distribution, which may be due to a trade-off between feeding, pre
dation avoidance, and endogenous fish requirements.