International Journal of Environment and Climate Change

This study developed and evaluated a machine-learning framework for forecasting mango yield in two mango-growing areas of Tamil Nadu, India: Periyakulam (Theni district) and Krishnagiri district. Annual rainfall, minimum and maximum temperature, relative humidity and mango-yield observations were compiled for 2010–2023 in Periyakulam and 2015–2023 in Krishnagiri. The resulting dataset included 114 district–variety–year observations covering six mango varieties. Descriptive assessments of inter-annual weather variability were combined with Pearson correlation analysis, analysis of variance (ANOVA) and a lightweight one-dimensional convolutional neural network (1D-CNN) regression model. Model performance was assessed using a 70:30 training–validation split and five-fold cross-validation, with Random Forest, Support Vector Regression and multiple linear regression as comparators. Yield differed significantly among varieties (F = 10.465, p < 0.001), whereas the district main effect was not significant (F = 0.406, p = 0.525). Validation root mean square error (RMSE) ranged from 3.12 to 3.94 kg/tree in Periyakulam and from 2.98 to 4.01 kg/tree in Krishnagiri. Across the two districts, validation correlations ranged from 88.65% to 95.64%. In the reported cross-validation comparison, the 1D-CNN yielded an RMSE of 3.42 kg/tree and a correlation of 91.8%, outperforming the baseline models. These findings suggest potential for weather-informed, district- and variety-specific mango-yield forecasting, although broader temporal and geographical validation is required.