Authors:
Bamory Ahmed Toru Koné
1
;
2
;
Rima Grati
3
;
1
;
Bassem Bouaziz
1
;
4
;
Khouloud Boukadi
2
;
1
and
Massimo Mecella
5
Affiliations:
1
University of Sfax, MIRACL Laboratory, Tunisia
;
2
Faculty of Economics and Management of Sfax, Tunisia
;
3
Computer Science Department, Zayed University, College of Technological Innovation, Abu Dhabi, U.A.E.
;
4
Higher Institute of Computer Science and Multimedia of Sfax, Tunisia
;
5
Dipartimento di Informatica e Sistemistica, Sapienza University, Rome, Italy
Keyword(s):
Precision Agriculture, Soil Moisture Prediction, Feature Selection, Explainable AI, Model Generalizability.
Abstract:
Accurate soil moisture prediction is critical for improving agricultural practices and managing water supplies. While feature selection techniques have proven useful in enhancing machine learning models’ performance in predicting soil moisture, their adaptation to different soil conditions remains limited. To address this gap, this study presents a novel multisite feature selection framework that draws on meteorological and soil data from three distinct locations with mineral, calcareous, and organic soils. The framework identifies soil-specific features through targeted selection processes and then uses SHAP, an explainable AI technique, to assess their global importance and influence. Furthermore, cross-site validation is performed to assess the transferability and generalizability of selected features, giving insight into their resilience across different environments. The proposed approach, which combines explainable AI and cross-site validation, provides a complete approach to u
nderstanding and improving feature relevance for soil moisture prediction. Overall, this study establishes the foundation for building more generalizable and robust predictive models, which will improve their applicability in a variety of agricultural and environmental scenarios.
(More)