NON-CLASSICAL APPROACHES TO THE STUDY OF THE ECOLOGY AND FISHERY OF Rastrineobola argentea (PELLEGRIN 1904) IN THE WINAM GULF OF LAKE VICTORIA, KENYA

Abstract

Many schools of thought tend to suggest that the central assumption in classical fisheries models may not necessarily hold and thus there is need to explore new approaches such as Bayesian Belief Networks (BBN), Artificial Neural Networks (ANN), Nominal and Ordinal Logistic Regression. This study used non classical methods such as logistic regression, Bayesian Belief Network (BBN), Artificial Neural Network (ANN) and Weibull/Lognormal distribution to study food habits, production and recruitment of R. argentea in Lake Victoria for the first time. Significant ontogenic changes in stomach content was determined for Thermocyclops oblingatus, Brachionus falcatus and Moina macrourus (p<0.0005) as compared to the baseline (Epiphanes spp.) for the 30-50, 50 and 30-50 mm length classes respectively. The odds ratio was 10.25-11.42 times for T. oblingatus and Moina macrurus as compared to Epiphanes. The BBN show that the Root Mean Square (RMS) change for Brachionus caudatus (0.00221), B. falcatus (0.00217), Epiphanes (0.00207), Keratella serrulata (0.00268), T. emini (0.00233), Bosmina longirostris (0.00217) and Daphnia lumholtzi (0.00258) and Trichocerca (0.00207) had the highest sensitivity of food items in the stomach as compared to the environment while B. calyciflorus, B. angularis and M. macrurus had the lowest sensitivity. Maximum Spawning Biomass (SB) and egg production was at a size between 40 and 60 mm TL. Egg production was best explained by a polynomial relationship of the fourth order with r2 of 0.959. Egg production, based on SB was significant for both Gamma and Weibull distribution (p<0.00005) according to the Shapiro-Wilks test. The location parameter was relatively consistent for both the Gamma (7,139) and Weibull (7,057) distributions, thereby providing similar recruitment threshold. Weibull distribution predicted a higher recruitment magnitude (scale parameter of 1,080,678) as compared to Gamma (354,600). The production modeling of R. argentea in Winam Gulf of Lake Victoria obtained the best ANN architecture of 10-9-1 based on environmental data and 12-6-1 based on fish catch statistics with 25 hidden layers and 30 hidden layers respectively, when the activation was based on the hyperbolic tangent function. Input importance analysis for environmental variables show that rainfall was the most significant variable (37%) followed by fisheries development classification (33%) and the lake level (17%) for environmental data. For fish catch statistics, the importance of fisheries development classification was 71.1%, Lates was 15.6%, Haplochromis was 6.6% and Bagrus was 4.2%. The actual catches versus output from the network had an average Absolute Error (AE) of 2,072 and 3,843 and an average Relative Absolute Error (RAE) of 14.2% and 20.7% for catch data and environmental data respectively. The ANN approach could be used to predict the catches of R. argentea in Lake Victoria during the different developmental stages of the fishery as well as projection of future production. Model data for both the environmental (r2 =0.852) and fish catches ((r2 =0.910) fitted well to the raw data. The non-classical methods offer robust alternatives for analysis of fisheries ecology data in light of data availability, nature of multispecies fishery and inadequacies of stock assessment models in tropical freshwater ecosystems. The study concludes that ordinal logistic regression best describes ontogenic changes in feeding while the BBN generated a stable feeding model for multiple food items. S-R relationship was best described by both Gamma and Weibull distributions for a given size at maturity, sex ratio, length-weight relationship and fecundity. The ANN consistently and adequately produced outputs that were consistent with target values from both environmental and catch data and could be used for predicting future values under varying fishing or environmental regime.