– Machine learning techniques are used for addressing challenges in soil fertility prediction problems in agriculture. The challenges include achieving accurate and timely predictions. To tackle these issues, a new technique called Fractional Proximity Normalized Imputation-based Improved Extreme Learning Machine (FPNI-IELM) is developed for soil fertility prediction with higher accuracy and minimal time consumption. The proposed FPNI-IELM technique consists of three major processes namely data acquisition, preprocessing, and feature engineering. Firstly, IoT devices are employed to collect soil health data for fertility prediction during the data acquisition phase. After data collection, preprocessing is carried out for accurate fertility prediction. In preprocessing, Fractional Gower’s proximity hot deck imputation is used for missing data detection. Subsequently, the maximum normalized residual test is applied for outlier detection. Following preprocessing, feature engineering is performed using the improved extreme learning machine to select significant features from the dataset, using Sokal–Michener’s simple matching to minimize the complexity of accurate soil fertility prediction. Experimental assessment is conducted on factors such as prediction accuracy, error rate, soil fertility prediction time, and space complexity. The quantitative assessment results reveal the effectiveness of the proposed technique, demonstrating higher prediction accuracy, lower error rates, and reduced time and space complexity compared to existing methods.