Artificial Neural Network Model for Predicting Post-Harvest Losses in Garri Production

Abstract

Idowu Olugbenga Adewumi, Oluwaseyifunmi Temitope Ajetunmobi, Adebayo James Durodola, Oluwasemipe Gladness Edward and Hameed Olalekan Olabiwonnu

In Nigeria, post-harvest losses during cassava processing are a significant issue, with national estimates indicating that 25–40% of harvested roots are lost before they are consumed or processed. This research created and assessed an artificial neural network (ANN) model to forecast losses in garri production based on a dataset of 16,892 samples gathered from various farms and processors in Ido Local Government, Ibadan, Oyo State Nigeria. Descriptive statistics showed an average cassava moisture content of 72.46% (SD = 4.33), an average lag between harvest and processing of 12.09 hours (SD = 6.62), a mean storage period of 7.55 days (SD = 4.06), and average projected losses of 32.43 kg (range: 5–60 kg). Baseline regression analysis (OLS) revealed that moisture (+0.2450 kg/unit, p < 0.001), harvest delay (+0.3187 kg/hr, p < 0.001), storage duration (+0.1891 kg/day, p < 0.001), and pest infestation (+3.5612 kg, p < 0.001) were significant factors in loss, whereas mechanical drying decreased losses by –2.2055 kg (p = 0.0061).

Three comparative models were employed: multiple linear regression (R2 = 0.78, MAE = 3.21, RMSE = 4.29), decision tree regression (R2 = 0.82, MAE = 2.75, RMSE = 3.91), and random forest regression (R2 = 0.89, MAE = 1.98, RMSE = 3.25). The ANN, set up with three hidden layers (64–32–16 neurons), ReLU activation, Adam optimizer (learning rate = 0.001), and 2,000 training epochs, attained outstanding performance (R2 = 0.92, MAE = 1.65, RMSE = 2.95, MSE = 8.72). Simulations based on scenarios demonstrated that mechanical drying alone decreased average expected losses from 9.60 kg to 7.40 kg (a 22.9% decrease), a 50% reduction in harvest delay cut losses to 6.20 kg (a 35.4% decrease), and a combined intervention reduced losses to 4.10 kg (a 57.3% decrease). Following deployment assessments showed a reduction in weekly losses from 40 kg to 15 kg (62.5% decrease), alongside an increase in farmer income from â?¦20,000 to â?¦28,000 (40% growth). These results affirm that ANN models exceed traditional statistical techniques and offer practical decision-making assistance for reducing losses. Combining predictive analytics with intervention design could decrease cassava post-harvest losses by more than 50% and boost farmer incomes by 30–40%, directly enhancing food security and rural livelihoods in sub-Saharan Africa.

PDF